diff --git a/blog/src/_data/media.json b/blog/src/_data/media.json index ae3ff34c6..adcd7e740 100644 --- a/blog/src/_data/media.json +++ b/blog/src/_data/media.json @@ -4610,5 +4610,16 @@ "status": "ready" }, "inline": {} + }, + "claude-haiku-5-5": { + "title": "Claude Haiku 5.5: price, benchmarks and where it fits", + "category": "concepts", + "hero": { + "file": "assets/media/claude-haiku-5-5/hero.png", + "alt": "Title card on white. Claude Haiku 5.5 in large ink type over a yellow underline, a Concepts kicker, the line Checked 8 Oct 2026, and a row of coloured dominoes part way through falling", + "prompt": "Still of the hero scene in assets/media/claude-haiku-5-5/motion.js.", + "status": "ready" + }, + "inline": {} } } diff --git a/blog/src/assets/media/claude-haiku-5-5/cups.png b/blog/src/assets/media/claude-haiku-5-5/cups.png new file mode 100644 index 000000000..51af3d5ba Binary files /dev/null and b/blog/src/assets/media/claude-haiku-5-5/cups.png differ diff --git a/blog/src/assets/media/claude-haiku-5-5/hero.png b/blog/src/assets/media/claude-haiku-5-5/hero.png new file mode 100644 index 000000000..8b7491b0e Binary files /dev/null and b/blog/src/assets/media/claude-haiku-5-5/hero.png differ diff --git a/blog/src/assets/media/claude-haiku-5-5/motion.css b/blog/src/assets/media/claude-haiku-5-5/motion.css new file mode 100644 index 000000000..a0101a107 --- /dev/null +++ b/blog/src/assets/media/claude-haiku-5-5/motion.css @@ -0,0 +1,13 @@ +/* Bright coded scenes for "Claude Haiku 5.5". Pairs with motion.js. */ +.prose figure.mg { margin: 42px 0; } +.prose figure.mg img, .mg .mg-svg { display: block; width: 100%; height: auto; border: 1px solid #1A1320; border-radius: 18px; background: #fff; } +.mg figcaption { font: 500 13px/1.5 "Space Grotesk", system-ui, sans-serif; color: #6B5878; padding-top: 10px; text-align: center; } +.mg .mg-svg { touch-action: manipulation; -webkit-tap-highlight-color: transparent; } +.mg .mg-svg text { -webkit-user-select: none; user-select: none; pointer-events: none; } +.mg .mg-hit { cursor: pointer; outline: none; } +.mg [role="slider"] { cursor: grab; touch-action: none; } +.mg .mg-ring { stroke-opacity: 0; transition: stroke-opacity 0.15s; } +.mg .mg-hit:focus-visible .mg-ring { stroke-opacity: 1; } +.mg [role="slider"]:focus-visible circle:nth-child(2) { stroke-width: 5; } +@media (hover: hover) { .mg .mg-hit:hover .mg-ring { stroke-opacity: 0.5; } } +@media (prefers-reduced-motion: reduce) { .mg .mg-ring { transition: none; } } diff --git a/blog/src/assets/media/claude-haiku-5-5/motion.js b/blog/src/assets/media/claude-haiku-5-5/motion.js new file mode 100644 index 000000000..bcbf7276c --- /dev/null +++ b/blog/src/assets/media/claude-haiku-5-5/motion.js @@ -0,0 +1,317 @@ +/* Bright coded scenes for "Claude Haiku 5.5". Vanilla JS, no library. +
holds a still image; this script swaps in a live SVG. + Plays only in view, honours reduced motion, "?mgstill" freezes each scene on its still frame. */ +(() => { + const NS = "http://www.w3.org/2000/svg"; + const INK = "#1A1320", YEL = "#FFCA54", BLUE = "#6C8EF5", LIL = "#B69CFF", MINT = "#7FD8AE", SKY = "#9AD8FF", SOFT = "#F1ECF9", MUTE = "#7A6A88"; + const SANS = '"Space Grotesk", system-ui, sans-serif', MONO = '"JetBrains Mono", ui-monospace, monospace'; + const el = (t, a = {}, p) => { const n = document.createElementNS(NS, t); for (const k in a) n.setAttribute(k, a[k]); if (p) p.appendChild(n); return n; }; + const txt = (p, x, y, s, a = {}) => { const n = el("text", Object.assign({ x, y, "font-family": SANS, "font-size": 24, "font-weight": 600, fill: INK }, a), p); n.textContent = s; return n; }; + const cl = (v, a = 0, b = 1) => Math.max(a, Math.min(b, v)); + const seg = (t, a, b) => cl((t - a) / (b - a)); + const oc = (x) => 1 - Math.pow(1 - x, 3); + const ob = (x) => { const c = 1.70158; return 1 + (c + 1) * Math.pow(x - 1, 3) + c * Math.pow(x - 1, 2); }; + const io = (x) => (x < 0.5 ? 4 * x * x * x : 1 - Math.pow(-2 * x + 2, 3) / 2); + const tf = (n, x, y, s = 1, r = 0) => n.setAttribute("transform", `translate(${x} ${y}) rotate(${r}) scale(${s})`); + const button = (g, label, fn) => { + g.setAttribute("role", "button"); g.setAttribute("tabindex", "0"); g.setAttribute("aria-label", label); g.classList.add("mg-hit"); + g.addEventListener("click", fn); + g.addEventListener("keydown", (e) => { if (e.key === "Enter" || e.key === " ") { e.preventDefault(); fn(); } }); + }; + const strip = (svg, y, w = 680) => { + el("rect", { x: (800 - w) / 2, y, width: w, height: 48, rx: 24, fill: SOFT }, svg); + return txt(svg, 400, y + 31, "", { "text-anchor": "middle", "font-size": 20, "font-weight": 500 }); + }; + + const GREY = "#B69CFF"; + const pillFit = (g, t, pad) => { const r = g.querySelector("rect"), w = t.getComputedTextLength() + pad; r.setAttribute("x", -w / 2); r.setAttribute("width", w); return w; }; + + const SC = {}; + + /* ---- 1. one stopwatch runs and stops; two paper labels carry the claim and its footnote ---- */ + SC.stopwatch = { + dur: 11, still: 8.6, + build(svg, api) { + txt(svg, 44, 62, "Anthropic's speed claim", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Announcement of 7 October 2026", { "font-size": 20, "font-weight": 500, fill: MUTE }); + txt(svg, 756, 60, "tap the stopwatch", { "text-anchor": "end", "font-size": 20, fill: MUTE, "font-weight": 500 }); + const CX = 236, CY = 318, R = 138; + const w = el("g", {}, svg); + const crown = el("g", {}, w); + el("rect", { x: -20, y: -R - 44, width: 40, height: 30, rx: 8, fill: SOFT, stroke: INK, "stroke-width": 3 }, crown); + el("rect", { x: -11, y: -R - 18, width: 22, height: 20, fill: "#fff", stroke: INK, "stroke-width": 3 }, w); + el("rect", { x: -13, y: -R - 30, width: 26, height: 22, rx: 6, fill: LIL, stroke: INK, "stroke-width": 3, transform: "rotate(42)" }, w); + el("circle", { r: R, fill: YEL, stroke: INK, "stroke-width": 3 }, w); + el("circle", { r: R - 22, fill: "#fff", stroke: INK, "stroke-width": 3 }, w); + for (let i = 0; i < 60; i++) { const big = i % 5 === 0; el("path", { d: `M0 ${-(R - 30)} V${-(R - (big ? 42 : 37))}`, stroke: big ? INK : GREY, "stroke-width": big ? 3 : 2, "stroke-linecap": "round", transform: `rotate(${i * 6})` }, w); } + const arc = el("circle", { r: R - 54, fill: "none", stroke: SKY, "stroke-width": 10, "stroke-linecap": "round", pathLength: 360, transform: "rotate(-90)" }, w); + txt(w, 0, 50, "Haiku 5.5", { "text-anchor": "middle", "font-size": 24, "font-weight": 700 }); + const hand = el("g", {}, w); + el("path", { d: `M0 22 V${-(R - 44)}`, stroke: INK, "stroke-width": 5, "stroke-linecap": "round" }, hand); + el("circle", { r: 10, fill: BLUE, stroke: INK, "stroke-width": 3 }, w); + el("circle", { r: R + 10, fill: "#fff", "fill-opacity": 0, stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, w); + button(w, "Stopwatch for Haiku 5.5. Press to run it again.", () => api.seek(0.2)); + const label = (y, fill, a, b, rot) => { + const g = el("g", {}, svg); + el("rect", { x: 0, y: -52, width: 300, height: 104, rx: 16, fill, stroke: INK, "stroke-width": 3 }, g); + el("circle", { cx: 26, cy: 0, r: 8, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + txt(g, 52, -8, a, { "font-size": 24, "font-weight": 700 }); + txt(g, 52, 24, b, { "font-size": 24, "font-weight": 500 }); + return { g, y, rot }; + }; + const A = label(252, YEL, "fastest to date", "at standard speed", -2.5), B = label(384, LIL, "Opus in Fast Mode", "is quicker", 2); + const cap = strip(svg, 506, 640); + return (t) => { + const run = oc(seg(t, 0.8, 3.6)), deg = 312 * run, out = 1 - seg(t, 10.4, 10.9); + hand.setAttribute("transform", `rotate(${deg * out})`); + arc.setAttribute("stroke-dasharray", `${Math.max(0.01, deg * out)} 400`); arc.setAttribute("opacity", deg * out < 2 ? 0 : 1); + const press = Math.sin(seg(t, 0.55, 0.85) * Math.PI) + Math.sin(seg(t, 3.55, 3.85) * Math.PI); + crown.setAttribute("transform", `translate(0 ${8 * press})`); + const jolt = Math.sin(seg(t, 3.6, 4.0) * Math.PI) * 0.03; + tf(w, CX, CY, 1 + jolt); + [[A, 4.2], [B, 5.4]].forEach(([l, s]) => { const k = ob(seg(t, s, s + 0.6)) * out; tf(l.g, 820 - 384 * k, l.y, 1, l.rot * k); l.g.setAttribute("opacity", k <= 0 ? 0 : 1); }); + cap.textContent = t < 5.4 || t > 10.4 ? "Anthropic calls Haiku 5.5 its fastest at standard speed" : "Anthropic's claim. We have not tried it."; + }; + }, + }; + + /* ---- 2. a small measuring cup fills, spills over its lip and carries on into a larger one ---- */ + SC.cups = { + dur: 12.5, still: 9.6, + build(svg) { + txt(svg, 44, 62, "One model, two price tiers", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Anthropic's prices for Haiku 5.5, per 1 million tokens", { "font-size": 20, "font-weight": 500, fill: MUTE }); + const G = 404; + const s1 = el("path", { stroke: SKY, "stroke-width": 12, fill: "none", "stroke-linecap": "round" }, svg); + const s2 = el("path", { stroke: SKY, "stroke-width": 12, fill: "none", "stroke-linecap": "round" }, svg); + /* tap */ + el("path", { d: "M50 150 H186 Q210 150 210 172 V184", stroke: INK, "stroke-width": 28, fill: "none", "stroke-linejoin": "round" }, svg); + el("path", { d: "M50 150 H186 Q210 150 210 172 V184", stroke: SOFT, "stroke-width": 22, fill: "none", "stroke-linejoin": "round" }, svg); + el("rect", { x: 44, y: 126, width: 14, height: 48, rx: 5, fill: LIL, stroke: INK, "stroke-width": 3 }, svg); + el("path", { d: "M128 136 V124", stroke: INK, "stroke-width": 3 }, svg); + const wh = el("rect", { x: 104, y: 112, width: 48, height: 13, rx: 6.5, fill: YEL, stroke: INK, "stroke-width": 3 }, svg); + const cup = (x, top, b, w, lip) => { + const r = 16, hw = lip ? 30 : 46, hh = lip ? 24 : 36; + const liq = el("path", { fill: SKY }, svg); + const n = lip ? 3 : 5; + for (let i = 1; i <= n; i++) el("path", { d: `M${x + 3} ${b - (i * (b - top)) / (n + 1)} h${i % 2 ? 18 : 28}`, stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round" }, svg); + el("path", { d: `M${x + w} ${top + 26} q${hw} 0 ${hw} ${hh} q0 ${hh} -${hw} ${hh}`, stroke: INK, "stroke-width": 3, fill: "none", "stroke-linecap": "round", transform: lip ? `translate(${2 * x + w} 0) scale(-1 1)` : "" }, svg); + el("path", { d: `M${x} ${top} V${b - r} Q${x} ${b} ${x + r} ${b} H${x + w - r} Q${x + w} ${b} ${x + w} ${b - r} V${top}` + (lip ? ` l14 -8` : ""), stroke: INK, "stroke-width": 3, fill: "none", "stroke-linecap": "round", "stroke-linejoin": "round" }, svg); + return (lv) => { const y = b - (b - top) * lv; liq.setAttribute("d", lv <= 0.01 ? "" : `M${x + 1.5} ${y} H${x + w - 1.5} V${b - r} Q${x + w - 1.5} ${b - 1.5} ${x + w - r} ${b - 1.5} H${x + r} Q${x + 1.5} ${b - 1.5} ${x + 1.5} ${b - r} Z`); return y; }; + }; + el("rect", { x: 122, y: 292, width: 176, height: G - 292, rx: 12, fill: SOFT, stroke: INK, "stroke-width": 3 }, svg); + const small = cup(146, 200, 292, 128, true), big = cup(420, 262, G, 270, false); + el("path", { d: `M44 ${G} H756`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + const card = (cx, fill, head, a, b) => { + const g = el("g", {}, svg); + el("rect", { x: -150, y: 0, width: 300, height: 128, rx: 16, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + el("path", { d: "M-150 42 H150 V16 Q150 0 134 0 H-134 Q-150 0 -150 16 Z", fill, stroke: INK, "stroke-width": 3, "stroke-linejoin": "round" }, g); + txt(g, 0, 29, head, { "text-anchor": "middle", "font-size": 20, "font-weight": 700 }); + [["input", a, 76], ["output", b, 110]].forEach(([k, v, y]) => { txt(g, -126, y, k, { "font-size": 22, "font-weight": 500 }); txt(g, 126, y, v, { "text-anchor": "end", "font-family": MONO, "font-size": 26, "font-weight": 700 }); }); + return { g, cx }; + }; + const c1 = card(210, YEL, "prompts up to 100k tokens", "$0.10", "$0.50"), c2 = card(570, LIL, "prompts over 100k tokens", "$0.50", "$2.50"); + return (t) => { + const out = 1 - seg(t, 11.6, 12.3); + const on = seg(t, 0.5, 0.8) * (1 - seg(t, 8.2, 8.5)); + const f1 = seg(t, 0.8, 3.4), f2 = seg(t, 4.0, 8.4); + const y1 = small(0.93 * f1 * out), y2 = big(0.72 * f2 * out); + wh.setAttribute("transform", `rotate(${-9 * (seg(t, 0.3, 0.6) - seg(t, 8.0, 8.3))} 128 118)`); + const wob = Math.sin(t * 20) * 1.2; + s1.setAttribute("d", `M210 188 V${y1 + 4}`); s1.setAttribute("opacity", on); s1.setAttribute("stroke-width", 12 + wob); + const over = seg(t, 3.5, 3.9) * (1 - seg(t, 8.5, 8.8)); + s2.setAttribute("d", `M284 198 Q${284 + 204 * over} 184 ${284 + 216 * over} ${198 + (y2 - 194) * over}`); s2.setAttribute("opacity", over <= 0 ? 0 : 1); s2.setAttribute("stroke-width", 12 - wob); + [[c1, 3.4], [c2, 5.2]].forEach(([c, s]) => { const k = ob(seg(t, s, s + 0.5)) * out; tf(c.g, c.cx, 432 + 20 * (1 - k), 0.7 + 0.3 * k); c.g.setAttribute("opacity", cl(k * 3)); }); + }; + }, + }; + + /* ---- 3. four paper planes, one benchmark at a time; each lands as far as its score ---- */ + SC.planes = { + dur: 14, still: 4.8, + build(svg, api) { + const M = [["Haiku 5.5", YEL], ["Haiku 4.5", SKY], ["GPT-6 Luna", LIL], ["Sonnet 5.5", BLUE]]; + const B = [ + ["GDPval-AA v2.1", ["1620", "735", "1437", "1840"], 2000, ""], + ["OSWorld 2.1, offline subset", ["72.4%", "15.7%", "48.9%", "83.9%"], 100, ""], + ["Humanity's Last Exam, no tools", ["45.9%", "10.2%", null, "56.9%"], 100, ""], + ["Terminal-Bench 4.0", ["39.2%", "0.0%", "16.4%", "70.6%"], 100, ""], + ["FrontierCode 1.1 (Main)", ["46.4%", null, "42.4%", "52.1%"], 100, "Xhigh effort"], + ]; + const N = B.length, P = 2.8, X0 = 236, LEN = 400, Y = [164, 232, 300, 368]; + const at = (b, i) => { const v = B[b][1][i]; return v == null ? { x: X0, nr: 1 } : { x: X0 + (parseFloat(v) / B[b][2]) * LEN, nr: 0 }; }; + txt(svg, 44, 62, "Anthropic's own figures", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Five rows we selected. We did not run them. NR means not reported.", { "font-size": 20, "font-weight": 500, fill: MUTE }); + el("path", { d: `M${X0 - 40} 132 V412`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + el("rect", { x: X0 - 40, y: 412, width: LEN + 64, height: 26, rx: 8, fill: SOFT, stroke: INK, "stroke-width": 3 }, svg); + for (let i = 1; i < 20; i++) el("path", { d: `M${X0 + (i * LEN) / 20} 412 v${i % 2 ? 9 : 15}`, stroke: INK, "stroke-width": 2.5 }, svg); + const pl = M.map(([name, col], i) => { + txt(svg, 44, Y[i] + 8, name, { "font-size": 22, "font-weight": 700 }); + const trail = el("path", { stroke: GREY, "stroke-width": 3, fill: "none", "stroke-linecap": "round", "stroke-dasharray": "2 9" }, svg); + const g = el("g", {}, svg); + const body = el("path", { d: "M-34 -15 L30 0 L-34 15 L-22 0 Z", fill: col, stroke: INK, "stroke-width": 3, "stroke-linejoin": "round" }, g); + el("path", { d: "M-22 0 H30", stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round" }, g); + const v = txt(svg, 0, Y[i] + 9, "", { "font-size": 24, "font-weight": 700 }); + const note = txt(svg, 0, Y[i] + 8, "", { "font-size": 20, "font-weight": 500, fill: MUTE }); + return { trail, g, body, v, note, col }; + }); + el("rect", { x: 172, y: 478, width: 456, height: 52, rx: 26, fill: SOFT }, svg); + const name = txt(svg, 400, 512, "", { "text-anchor": "middle", "font-size": 22, "font-weight": 700 }); + let user = null, shown = 0; + const step = (d) => { user = { from: shown, to: (shown + d + N) % N, t0: api.now() }; api.poke(); }; + [[-1, 130, "Previous benchmark", "M5 -9 L-5 0 L5 9"], [1, 670, "Next benchmark", "M-5 -9 L5 0 L-5 9"]].forEach(([d, x, label, arrow]) => { + const g = el("g", { transform: `translate(${x} 504)` }, svg); + el("circle", { r: 32, fill: "none", stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, g); + el("circle", { r: 26, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + el("path", { d: arrow, stroke: INK, "stroke-width": 3.5, fill: "none", "stroke-linecap": "round", "stroke-linejoin": "round" }, g); + button(g, label, () => step(d)); + g.addEventListener("keydown", (e) => { const s = e.key === "ArrowRight" || e.key === "ArrowUp" ? 1 : e.key === "ArrowLeft" || e.key === "ArrowDown" ? -1 : 0; if (s) { e.preventDefault(); step(s); } }); + }); + return (t, now) => { + let idx, prev, u; + if (user) { idx = user.to; prev = user.from; u = cl((now - user.t0) / P); } + else { idx = Math.floor(t / P) % N; prev = (idx + N - 1) % N; u = (t % P) / P; } + shown = idx; name.textContent = B[idx][0]; + pl.forEach((p, i) => { + const a = at(prev, i), c = at(idx, i), k = io(seg(u, 0.02 + i * 0.03, 0.36 + i * 0.03)), x = a.x + (c.x - a.x) * k, lift = Math.sin(k * Math.PI); + tf(p.g, x, Y[i] - 16 * lift, 1, -9 * lift * Math.sign(c.x - a.x || 1) * Math.cos(k * Math.PI)); + p.trail.setAttribute("d", x - X0 < 34 ? "" : `M${X0 - 24} ${Y[i]} H${x - 46}`); + const s = B[idx][1][i], show = seg(u, 0.4 + i * 0.03, 0.5 + i * 0.03); + p.v.textContent = s == null ? "NR" : s; p.v.setAttribute("x", c.x + 44); p.v.setAttribute("opacity", show); + const nt = i === 3 ? B[idx][3] : ""; p.note.textContent = nt; if (nt) { p.note.setAttribute("x", c.x + 44 + p.v.getComputedTextLength() + 12); p.note.setAttribute("opacity", show); } + p.body.setAttribute("fill", c.nr && k > 0.5 ? "#fff" : p.col); p.body.setAttribute("stroke-dasharray", c.nr && k > 0.5 ? "6 6" : "none"); + }); + }; + }, + }; + + /* ---- 4. a postcard rack turns; the card at the front flips to show what the alias gives there ---- */ + SC.postcards = { + dur: 14.6, still: 12.6, + build(svg, api) { + const C = [[["Claude Platform", "on AWS"], "Haiku 4.5", SKY, "Claude Platform on AWS"], [["Amazon", "Bedrock"], "Haiku 4.5", LIL, "Amazon Bedrock"], [["Google Cloud's", "Agent Platform"], "Haiku 4.5", YEL, "Google Cloud's Agent Platform"], [["Microsoft", "Foundry"], "Haiku 4.5", BLUE, "Microsoft Foundry"], [["Anthropic", "API"], "Haiku 5.5", MINT, "the Anthropic API"]]; + const N = C.length, P = 2.6, R = 262, CY = 296, W = 230, H = 276; + txt(svg, 44, 62, "Where the haiku alias points", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Claude Code model docs, read 8 Oct 2026", { "font-size": 20, "font-weight": 500, fill: MUTE }); + txt(svg, 756, 60, "tap to turn", { "text-anchor": "end", "font-size": 20, fill: MUTE, "font-weight": 500 }); + el("ellipse", { cx: 400, cy: CY, rx: R, ry: 22, fill: "none", stroke: GREY, "stroke-width": 3, "stroke-dasharray": "2 9", "stroke-linecap": "round" }, svg); + const layer = el("g", {}, svg); + el("rect", { x: 34, y: 116, width: 732, height: 374, rx: 26, fill: "#fff", "fill-opacity": 0, stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, layer); + const pole = el("g", {}, layer); + el("rect", { x: 330, y: 462, width: 140, height: 20, rx: 10, fill: SOFT, stroke: INK, "stroke-width": 3 }, pole); + el("path", { d: "M400 128 V462", stroke: INK, "stroke-width": 6, "stroke-linecap": "round" }, pole); + el("circle", { cx: 400, cy: 124, r: 11, fill: YEL, stroke: INK, "stroke-width": 3 }, pole); + const cards = C.map(([nm, model, col]) => { + const g = el("g", {}, layer), flip = el("g", {}, g); + const front = el("g", {}, flip), back = el("g", {}, flip); + el("rect", { x: -W / 2, y: -H / 2, width: W, height: H, rx: 18, fill: "#fff", stroke: INK, "stroke-width": 3 }, front); + const ft = el("g", {}, front); + el("rect", { x: W / 2 - 62, y: -H / 2 + 18, width: 44, height: 52, rx: 6, fill: col, stroke: INK, "stroke-width": 3 }, ft); + nm.forEach((s, i) => txt(ft, 0, -8 + i * 34, s, { "text-anchor": "middle", "font-size": 26, "font-weight": 700 })); + el("path", { d: `M${-W / 2 + 28} 74 H${W / 2 - 28} M${-W / 2 + 28} 98 H${W / 2 - 70}`, stroke: SOFT, "stroke-width": 6, "stroke-linecap": "round" }, ft); + el("rect", { x: -W / 2, y: -H / 2, width: W, height: H, rx: 18, fill: model === "Haiku 5.5" ? MINT : SOFT, stroke: INK, "stroke-width": 3 }, back); + const bt = el("g", {}, back); + const pg = el("g", { transform: "translate(0 -92)" }, bt); + el("rect", { x: -52, y: -22, width: 104, height: 44, rx: 12, fill: "#fff", stroke: INK, "stroke-width": 3 }, pg); + txt(pg, 0, 8, "haiku", { "text-anchor": "middle", "font-family": MONO, "font-size": 24, "font-weight": 700 }); + txt(bt, 0, -34, "resolves to", { "text-anchor": "middle", "font-size": 24, "font-weight": 500 }); + txt(bt, 0, 14, model, { "text-anchor": "middle", "font-size": 40, "font-weight": 700 }); + el("path", { d: `M${-W / 2 + 28} 40 H${W / 2 - 28}`, stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round", "stroke-dasharray": "2 8" }, bt); + nm.forEach((s, i) => txt(bt, 0, 78 + i * 30, s, { "text-anchor": "middle", "font-size": 24, "font-weight": 600 })); + return { g, flip, front, back, ft, bt }; + }); + const cap = strip(svg, 506, 700); + let last = 0, key = ""; + button(layer, "Postcard rack. Press to turn to the next provider.", () => { const k = Math.min(N - 1, Math.floor(last / P)); api.seek(((k + 1) % N) * P + 0.02); }); + return (t) => { + last = t; + const k = Math.min(N - 1, Math.floor(t / P)), u = t - k * P, pos = k - 1 + io(seg(u, 0, 0.9)), back = seg(t, N * P + 0.9, N * P + 1.4); + const order = [[0, pole]]; + cards.forEach((c, i) => { + const a = ((i - pos) * 2 * Math.PI) / N, z = Math.cos(a), s = 0.64 + 0.36 * (z + 1) / 2; + tf(c.g, 400 + R * Math.sin(a), CY + 16 * z, s); + const f = (i < k ? 1 : i === k ? io(seg(u, 1.05, 1.6)) : 0) * (1 - io(back)), sx = Math.abs(Math.cos(f * Math.PI)); + c.flip.setAttribute("transform", `scale(${Math.max(0.02, sx)} 1)`); + c.front.setAttribute("display", f > 0.5 ? "none" : ""); c.back.setAttribute("display", f > 0.5 ? "" : "none"); + const vis = seg(z, -0.05, 0.28); c.ft.setAttribute("opacity", vis); c.bt.setAttribute("opacity", vis); + order.push([z, c.g]); + }); + order.sort((p, q) => p[0] - q[0]); + const nk = order.map((o) => (o[1] === pole ? "p" : cards.findIndex((c) => c.g === o[1]))).join(""); + if (nk !== key) { key = nk; order.forEach((o) => layer.appendChild(o[1])); } + cap.textContent = u < 1.35 || back > 0 ? "In Claude Code, the haiku alias depends on your provider" : `On ${C[k][3]}, haiku resolves to ${C[k][1]}`; + }; + }, + }; + + /* ---- title card, 16 by 9: a row of dominoes part way through falling ---- */ + SC.hero = { + w: 960, h: 540, dur: 9, still: 2.7, + build(svg) { + const kg = el("g", { transform: "translate(64 62)" }, svg); + const kr = el("rect", { x: 0, y: 0, height: 40, rx: 20, fill: SOFT, stroke: INK, "stroke-width": 2.5 }, kg); + const kt = txt(kg, 20, 28, "Concepts", { "font-size": 22, "font-weight": 700 }); + kr.setAttribute("width", kt.getComputedTextLength() + 40); + const title = txt(svg, 62, 206, "Claude Haiku 5.5", { "font-size": 96, "font-weight": 700 }); + el("path", { d: `M66 236 H${62 + title.getComputedTextLength() - 6}`, stroke: YEL, "stroke-width": 14, "stroke-linecap": "round" }, svg); + txt(svg, 64, 292, "Checked 8 Oct 2026", { "font-size": 26, "font-weight": 500, fill: MUTE }); + const G = 478, DW = 20, DH = 96, D = 52, n = 15, TH = (Math.acos(DW / D) * 180) / Math.PI; + el("path", { d: `M64 ${G} H896`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + const COL = [YEL, SKY, LIL, MINT, BLUE]; + const ds = Array.from({ length: n }, (_, i) => { + const g = el("g", {}, svg); + el("rect", { x: -DW, y: -DH, width: DW, height: DH, rx: 5, fill: COL[i % 5], stroke: INK, "stroke-width": 3 }, g); + return g; + }); + const RAD = Math.PI / 180; + const fits = (a, b) => { + const px = DH * Math.sin(a), py = DH * Math.cos(a), bx = D - DW * Math.cos(b), by = DW * Math.sin(b); + const along = (px - bx) * Math.sin(b) + (py - by) * Math.cos(b), side = (px - bx) * Math.cos(b) - (py - by) * Math.sin(b); + if (along <= DH) return side <= 0.3; + return (bx + DH * Math.sin(b)) * Math.cos(a) - (by + DH * Math.cos(b)) * Math.sin(a) >= -0.3; + }; + return (t) => { + const up = 1 - io(seg(t, 7.6, 8.6)); + let next = null; + for (let i = n - 1; i >= 0; i--) { + const k = seg(t, 0.6 + i * 0.24, 1.3 + i * 0.24); + let a = 90 * k * k * RAD; + if (next != null && !fits(a, next)) { let lo = 0, hi = a; for (let j = 0; j < 22; j++) { const m = (lo + hi) / 2; if (fits(m, next)) lo = m; else hi = m; } a = lo; } + next = a; + ds[i].setAttribute("transform", `translate(${104 + i * D} ${G}) rotate(${(a / RAD) * up})`); + } + }; + }, + }; + + /* ---- player ---- */ + const q = new URLSearchParams(location.search), frozen = q.has("mgstill"), fixed = parseFloat(q.get("mgstill")); + const calm = matchMedia("(prefers-reduced-motion: reduce)").matches; + const mount = (fig) => { + const sc = SC[fig.dataset.scene], img = fig.querySelector("img"); if (!sc || !img) return; + const W = sc.w || 800, H = sc.h || 600; + const svg = el("svg", { viewBox: `0 0 ${W} ${H}`, class: "mg-svg", role: "group", "aria-label": img.alt || "" }); + svg.style.aspectRatio = `${W} / ${H}`; + el("rect", { width: W, height: H, fill: "#fff" }, svg); + img.replaceWith(svg); + let off = 0, raf = 0, seen = false, held = calm; + const base = performance.now(), clock = () => (performance.now() - base) / 1000; + let draw = () => {}; + const frame = () => { const now = clock(); draw(held ? sc.still : (((now + off) % sc.dur) + sc.dur) % sc.dur, now); }; + const tick = () => { cancelAnimationFrame(raf); frame(); if (seen && !held) raf = requestAnimationFrame(tick); }; + const api = { now: clock, seek: (s) => { off = s - clock(); held = false; tick(); }, poke: () => { held = false; tick(); } }; + draw = sc.build(svg, api); + if (frozen) { draw(isNaN(fixed) ? sc.still : fixed, 0); return; } + new IntersectionObserver(([e]) => { + seen = e.isIntersecting; + if (seen) { tick(); if (!fig.dataset.sent && window.posthog) { fig.dataset.sent = 1; window.posthog.capture("blog_scene_view", { slug: location.pathname.split("/").filter(Boolean).pop(), scene: fig.dataset.scene }); } } + else cancelAnimationFrame(raf); + }, { threshold: 0.25 }).observe(fig); + svg.addEventListener("click", () => { if (!fig.dataset.hit && window.posthog) { fig.dataset.hit = 1; window.posthog.capture("blog_scene_interact", { slug: location.pathname.split("/").filter(Boolean).pop(), scene: fig.dataset.scene }); } }); + frame(); + }; + const go = () => document.querySelectorAll("figure.mg[data-scene]").forEach(mount); + const fonts = document.fonts && document.fonts.load ? Promise.all([document.fonts.load('700 32px "Space Grotesk"'), document.fonts.load('500 20px "Space Grotesk"'), document.fonts.load('700 20px "JetBrains Mono"')]).catch(() => {}) : Promise.resolve(); + const start = () => fonts.then(go, go); + if (document.readyState === "loading") document.addEventListener("DOMContentLoaded", start); else start(); +})(); diff --git a/blog/src/assets/media/claude-haiku-5-5/planes.png b/blog/src/assets/media/claude-haiku-5-5/planes.png new file mode 100644 index 000000000..b57c59d02 Binary files /dev/null and b/blog/src/assets/media/claude-haiku-5-5/planes.png differ diff --git a/blog/src/assets/media/claude-haiku-5-5/postcards.png b/blog/src/assets/media/claude-haiku-5-5/postcards.png new file mode 100644 index 000000000..24b9bcac6 Binary files /dev/null and b/blog/src/assets/media/claude-haiku-5-5/postcards.png differ diff --git a/blog/src/assets/media/claude-haiku-5-5/stopwatch.png b/blog/src/assets/media/claude-haiku-5-5/stopwatch.png new file mode 100644 index 000000000..933b39827 Binary files /dev/null and b/blog/src/assets/media/claude-haiku-5-5/stopwatch.png differ diff --git a/blog/src/posts/claude-code-subagents-vs-multi-agent-harness.md b/blog/src/posts/claude-code-subagents-vs-multi-agent-harness.md index f6051307c..019e31a5c 100644 --- a/blog/src/posts/claude-code-subagents-vs-multi-agent-harness.md +++ b/blog/src/posts/claude-code-subagents-vs-multi-agent-harness.md @@ -79,7 +79,7 @@ You can also force it: name the subagent in plain words, @-mention it with `@age ## How do you set a subagent's model and tools? -Use `model` and `tools` in the front matter. The model is resolved in this order: a model Claude passes when it spawns the subagent, then the `model` field, then the `CLAUDE_CODE_SUBAGENT_MODEL` variable, then your session's model. Since 2.1.251 that variable is only a default. To force one model on every subagent, also set `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1` (2.1.257 and later). Run `/tasks` to see which model each subagent is on. +Use `model` and `tools` in the front matter. The model is resolved in this order: a model Claude passes when it spawns the subagent, then the `model` field, then the `CLAUDE_CODE_SUBAGENT_MODEL` variable, then your session's model. Since 2.1.251 that variable is only a default. To force one model on every subagent, also set `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1` (2.1.257 and later). Run `/tasks` to see which model each subagent is on. The newest Haiku is [Claude Haiku 5.5](/blog/claude-haiku-5-5/). `tools` is an allowlist and `disallowedTools` a denylist. A `disallowedTools` entry such as `Bash(git push *)` removes the whole Bash tool, so block single commands with a deny rule in settings instead. The Task tool was renamed Agent in 2.1.63, and old `Task(...)` rules still work. diff --git a/blog/src/posts/claude-haiku-5-5.md b/blog/src/posts/claude-haiku-5-5.md new file mode 100644 index 000000000..560779a77 --- /dev/null +++ b/blog/src/posts/claude-haiku-5-5.md @@ -0,0 +1,112 @@ +--- +title: "Claude Haiku 5.5: price, benchmarks and where it fits" +description: "Claude Haiku 5.5 is Anthropic's new small model. Its tiered price, Anthropic's own scores and how to pick it in Claude Code, checked 8 Oct 2026." +date: 2026-10-08 +category: concepts +categoryLabel: Concepts +type: Non-technical +primaryKeyword: "claude haiku 5.5" +secondaryKeywords: ["claude haiku", "haiku 5.5", "claude haiku 5.5 pricing", "claude haiku 5.5 vs haiku 4.5", "claude haiku 5.5 claude code", "claude-haiku-5-5"] +tags: ["Concepts", "Claude Code", "AI Agents"] +faq: + - q: "How much does Claude Haiku 5.5 cost?" + a: "Anthropic's announcement of 7 October 2026 lists $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 input and $2.50 output for prompts over 100,000 tokens. Cache reads are $0.01 and $0.05. Checked 8 Oct 2026." + - q: "Is Claude Haiku 5.5 better than Haiku 4.5?" + a: "On Anthropic's own figures, yes on every row where both have a score. For example, the announcement lists 72.4% for Haiku 5.5 and 15.7% for Haiku 4.5 on the offline subset of OSWorld 2.1. We did not run these tests." + - q: "Is Claude Haiku 5.5 better than Sonnet 5.5?" + a: "No. Sonnet 5.5 is ahead on every row of Anthropic's own benchmark table, and the announcement says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks. Haiku 5.5 is the cheaper and faster option." + - q: "What is the model ID for Claude Haiku 5.5?" + a: "It is claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock, according to Anthropic's model page read on 8 Oct 2026." + - q: "How do I use Claude Haiku 5.5 in Claude Code?" + a: "Claude Code's model docs say to run /model claude-haiku-5-5 in a session, or start with claude --model claude-haiku-5-5. The same page says to use Claude Code v2.1.293 or later with Haiku 5.5. Read on 8 Oct 2026." +--- + +Claude Haiku 5.5 is Anthropic's new small model: much cheaper than Haiku 4.5 on prompts up to 100,000 tokens, and behind Sonnet 5.5 on every row of Anthropic's own benchmark table. Anthropic released it on 7 October 2026. We have not tried it. Checked 8 Oct 2026. + +[Munder Difflin](https://harnessmd.com/download) is free and open source: a desktop app that runs a team of coding agents such as Claude Code, Codex and Gemini CLI on your own computer. A cheaper small model matters when several agents run at once, which is why we cover it. The [install guide](/blog/how-to-install-and-use-munder-difflin/) covers setup and the [Concepts hub](/blog/topics/concepts/) explains the terms. + +## What is Claude Haiku 5.5? + +It is Anthropic's new small model, with the model ID `claude-haiku-5-5`. Anthropic's [announcement](https://www.anthropic.com/claude-haiku-5-5), dated October 7, 2026, calls it "the cheapest, fastest, and most capable small model we've ever released" and says it is designed for high-volume, cost-sensitive tasks. + +The speed claim has a footnote. Anthropic says Haiku 5.5 is its fastest model to date at each model's standard speed, although it runs less quickly than its Opus models in Fast Mode. + +Anthropic's [model page](https://platform.claude.com/docs/en/models/haiku-5-5/overview), read on 8 Oct 2026, lists a 1M token context window and a maximum output of 128K tokens. The announcement adds that this is the first Haiku-class model with an adjustable effort setting. + +The [Hacker News thread](https://news.ycombinator.com/item?id=49996437) had more than 300 points and 140 comments when we checked on 8 Oct 2026. + +
Animation. A stopwatch labelled Haiku 5.5 starts, its hand sweeps round and stops. Two paper labels slide in beside it: one reads fastest to date at standard speed, the other reads Opus in Fast Mode is quicker.
Anthropic calls Haiku 5.5 its fastest model to date at standard speed, with Opus in Fast Mode quicker. Announcement of 7 October 2026.
+ +## What does Claude Haiku 5.5 cost? + +It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, as of 8 Oct 2026, and five times that above. These are the prices in the table in Anthropic's [announcement](https://www.anthropic.com/claude-haiku-5-5), per 1 million tokens. + +| Price per 1M tokens | Haiku 5.5, prompts up to 100k | Haiku 5.5, prompts over 100k | Haiku 4.5 | Sonnet 5.5 | +| --- | --- | --- | --- | --- | +| Input | $0.10 | $0.50 | $1.00 | $2.00 | +| Output | $0.50 | $2.50 | $5.00 | $10.00 | +| Cache writes | $0.125 | $0.625 | $1.25 | $2.50 | +| Cache reads | $0.01 | $0.05 | $0.10 | $0.10 | + +Anthropic says prompts up to 100,000 tokens made up around 90% of requests to its previous Haiku model. Its headline is that Haiku 5.5 costs around 75% less to run on average. A footnote explains the gap: the price is 90% lower for requests up to 100,000 tokens, less for larger ones, and an updated tokenizer uses slightly more tokens per task. The model page says the same text counts as approximately 30% more tokens than on Haiku 4.5. + +The model page also lists a 50% Batch API discount on input and output. For the wider bill, see [how much Claude Code costs](/blog/how-much-does-claude-code-cost/). + +
Animation. A tap fills a small measuring cup, and a card under it headed prompts up to 100k tokens shows $0.10 input and $0.50 output. The water spills over the lip into a larger cup, and its card headed prompts over 100k tokens shows $0.50 input and $2.50 output.
Anthropic's prices per 1 million tokens, in two tiers by prompt size. Checked 8 Oct 2026.
+ +## How does Claude Haiku 5.5 score on benchmarks? + +On Anthropic's own figures it is far ahead of Haiku 4.5 and behind Sonnet 5.5. The five rows below are our selection from the eight rows in the [announcement](https://www.anthropic.com/claude-haiku-5-5) table. We did not run them. NR means not reported. + +| Benchmark (as named by Anthropic) | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 | +| --- | --- | --- | --- | --- | +| GDPval-AA v2.1 | 1620 | 735 | 1437 | 1840 | +| OSWorld 2.1, offline subset | 72.4% | 15.7% | 48.9% | 83.9% | +| Humanity's Last Exam, no tools | 45.9% | 10.2% | NR | 56.9% | +| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% | +| FrontierCode 1.1 (Main) | 46.4% | NR | 42.4% | 52.1% | + +Sonnet 5.5 leads on all eight rows of the full table. Anthropic marks its FrontierCode figure for Sonnet 5.5 as Xhigh effort. Haiku 5.5 is ahead of GPT-6 Luna on all six rows where Anthropic reports both. + +The coding gap is the one to notice. On Terminal-Bench 4.0, Haiku 5.5 is at 39.2% against 70.6% for Sonnet 5.5. Anthropic says so itself: "Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0." + +
Animation. Four paper planes for Haiku 5.5, Haiku 4.5, GPT-6 Luna and Sonnet 5.5 glide along a ruled strip and land at distances set by their scores on one benchmark at a time. Arrow buttons step through the five benchmarks.
Anthropic's own figures, five rows we selected from its table of 7 October 2026. We did not run them.
+ +## Where can you use Claude Haiku 5.5? + +Anthropic says it "is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure". The [model page](https://platform.claude.com/docs/en/models/haiku-5-5/overview) lists the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. The ID is `claude-haiku-5-5`, except on Amazon Bedrock, where it is `anthropic.claude-haiku-5-5`. + +Neither page says which Claude app plans include it, so we leave that out. + +## How do you use Claude Haiku 5.5 in Claude Code? + +Run `/model claude-haiku-5-5` in a session, or start with `claude --model claude-haiku-5-5`. That wording is from Claude Code's [model configuration docs](https://code.claude.com/docs/en/model-config), read on 8 Oct 2026. The same page says to use v2.1.293 or later with Haiku 5.5. + +Three details from the docs: + +* **The alias depends on your provider.** `haiku` resolves to Haiku 5.5 on the Anthropic API. On Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry it resolves to Haiku 4.5. +* **Long prompts cost more.** The docs repeat that a Haiku 5.5 request costs more per token when its prompt is longer than 100K tokens. +* **Subagents take a model.** The [subagent docs](https://code.claude.com/docs/en/sub-agents) say the `model` field accepts an alias such as `haiku`, a full model ID or `inherit`. `CLAUDE_CODE_SUBAGENT_MODEL` sets a default. + +Our post on [Claude Code subagents](/blog/claude-code-subagents-vs-multi-agent-harness/) walks through that file. + +
Animation. A postcard rack turns slowly. Each card names a provider, and as it faces front it flips to show which model the haiku alias gives: Haiku 5.5 on the Anthropic API card, Haiku 4.5 on the other cards.
In Claude Code the haiku alias resolves to Haiku 5.5 on the Anthropic API and to Haiku 4.5 on the other listed providers. Docs read 8 Oct 2026.
+ +## What does it mean if you run a team of coding agents? + +We think it makes the split between a planner and its helpers cheaper to try. Anthropic says Haiku 5.5 "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work", and that it suits narrowly scoped tasks like compaction, summarization or subagent work. + +Two customer quotes from the announcement, the first on subagents and the second on speed. Rogo: "The short and high-volume work is where Claude Haiku 5.5 fits for us, like quick lookups, subagents, and summaries." Asana: "Compared with the model we use today, we saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn." + +Our view: put a bigger model on planning and hard debugging, and a small fast one on searching, summarising and running tests. The small one fetches, the big one decides. This is an idea, not a test. We have not tried Haiku 5.5. Our piece on [model routing](/blog/do-more-with-less-model-routing/) explains the reasoning. + +## What should you do now? + +Our view: try it on routine work first, and keep your bigger model where the task is hard. + +1. Update Claude Code to v2.1.293 or later. +2. Set `model: haiku` on one subagent that only searches or summarises. +3. Watch prompt size. The lower price applies up to 100,000 tokens. +4. Compare results on your own repo before trusting any table. + + diff --git a/docs/blog/agent-tools-today-2026-10-07/index.html b/docs/blog/agent-tools-today-2026-10-07/index.html index 0e2547c2c..0ed252b1d 100644 --- a/docs/blog/agent-tools-today-2026-10-07/index.html +++ b/docs/blog/agent-tools-today-2026-10-07/index.html @@ -241,7 +241,7 @@

How this list is built
- +

Related in News

diff --git a/docs/blog/agents-rule-of-two/index.html b/docs/blog/agents-rule-of-two/index.html index 5934a60d6..0e4bc745f 100644 --- a/docs/blog/agents-rule-of-two/index.html +++ b/docs/blog/agents-rule-of-two/index.html @@ -331,50 +331,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/ai-agent-glossary/index.html b/docs/blog/ai-agent-glossary/index.html index 4f283f457..9e24912ad 100644 --- a/docs/blog/ai-agent-glossary/index.html +++ b/docs/blog/ai-agent-glossary/index.html @@ -303,50 +303,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/assets/media/claude-haiku-5-5/cups.png b/docs/blog/assets/media/claude-haiku-5-5/cups.png new file mode 100644 index 000000000..51af3d5ba Binary files /dev/null and b/docs/blog/assets/media/claude-haiku-5-5/cups.png differ diff --git a/docs/blog/assets/media/claude-haiku-5-5/hero.png b/docs/blog/assets/media/claude-haiku-5-5/hero.png new file mode 100644 index 000000000..8b7491b0e Binary files /dev/null and b/docs/blog/assets/media/claude-haiku-5-5/hero.png differ diff --git a/docs/blog/assets/media/claude-haiku-5-5/motion.css b/docs/blog/assets/media/claude-haiku-5-5/motion.css new file mode 100644 index 000000000..a0101a107 --- /dev/null +++ b/docs/blog/assets/media/claude-haiku-5-5/motion.css @@ -0,0 +1,13 @@ +/* Bright coded scenes for "Claude Haiku 5.5". Pairs with motion.js. */ +.prose figure.mg { margin: 42px 0; } +.prose figure.mg img, .mg .mg-svg { display: block; width: 100%; height: auto; border: 1px solid #1A1320; border-radius: 18px; background: #fff; } +.mg figcaption { font: 500 13px/1.5 "Space Grotesk", system-ui, sans-serif; color: #6B5878; padding-top: 10px; text-align: center; } +.mg .mg-svg { touch-action: manipulation; -webkit-tap-highlight-color: transparent; } +.mg .mg-svg text { -webkit-user-select: none; user-select: none; pointer-events: none; } +.mg .mg-hit { cursor: pointer; outline: none; } +.mg [role="slider"] { cursor: grab; touch-action: none; } +.mg .mg-ring { stroke-opacity: 0; transition: stroke-opacity 0.15s; } +.mg .mg-hit:focus-visible .mg-ring { stroke-opacity: 1; } +.mg [role="slider"]:focus-visible circle:nth-child(2) { stroke-width: 5; } +@media (hover: hover) { .mg .mg-hit:hover .mg-ring { stroke-opacity: 0.5; } } +@media (prefers-reduced-motion: reduce) { .mg .mg-ring { transition: none; } } diff --git a/docs/blog/assets/media/claude-haiku-5-5/motion.js b/docs/blog/assets/media/claude-haiku-5-5/motion.js new file mode 100644 index 000000000..bcbf7276c --- /dev/null +++ b/docs/blog/assets/media/claude-haiku-5-5/motion.js @@ -0,0 +1,317 @@ +/* Bright coded scenes for "Claude Haiku 5.5". Vanilla JS, no library. +
holds a still image; this script swaps in a live SVG. + Plays only in view, honours reduced motion, "?mgstill" freezes each scene on its still frame. */ +(() => { + const NS = "http://www.w3.org/2000/svg"; + const INK = "#1A1320", YEL = "#FFCA54", BLUE = "#6C8EF5", LIL = "#B69CFF", MINT = "#7FD8AE", SKY = "#9AD8FF", SOFT = "#F1ECF9", MUTE = "#7A6A88"; + const SANS = '"Space Grotesk", system-ui, sans-serif', MONO = '"JetBrains Mono", ui-monospace, monospace'; + const el = (t, a = {}, p) => { const n = document.createElementNS(NS, t); for (const k in a) n.setAttribute(k, a[k]); if (p) p.appendChild(n); return n; }; + const txt = (p, x, y, s, a = {}) => { const n = el("text", Object.assign({ x, y, "font-family": SANS, "font-size": 24, "font-weight": 600, fill: INK }, a), p); n.textContent = s; return n; }; + const cl = (v, a = 0, b = 1) => Math.max(a, Math.min(b, v)); + const seg = (t, a, b) => cl((t - a) / (b - a)); + const oc = (x) => 1 - Math.pow(1 - x, 3); + const ob = (x) => { const c = 1.70158; return 1 + (c + 1) * Math.pow(x - 1, 3) + c * Math.pow(x - 1, 2); }; + const io = (x) => (x < 0.5 ? 4 * x * x * x : 1 - Math.pow(-2 * x + 2, 3) / 2); + const tf = (n, x, y, s = 1, r = 0) => n.setAttribute("transform", `translate(${x} ${y}) rotate(${r}) scale(${s})`); + const button = (g, label, fn) => { + g.setAttribute("role", "button"); g.setAttribute("tabindex", "0"); g.setAttribute("aria-label", label); g.classList.add("mg-hit"); + g.addEventListener("click", fn); + g.addEventListener("keydown", (e) => { if (e.key === "Enter" || e.key === " ") { e.preventDefault(); fn(); } }); + }; + const strip = (svg, y, w = 680) => { + el("rect", { x: (800 - w) / 2, y, width: w, height: 48, rx: 24, fill: SOFT }, svg); + return txt(svg, 400, y + 31, "", { "text-anchor": "middle", "font-size": 20, "font-weight": 500 }); + }; + + const GREY = "#B69CFF"; + const pillFit = (g, t, pad) => { const r = g.querySelector("rect"), w = t.getComputedTextLength() + pad; r.setAttribute("x", -w / 2); r.setAttribute("width", w); return w; }; + + const SC = {}; + + /* ---- 1. one stopwatch runs and stops; two paper labels carry the claim and its footnote ---- */ + SC.stopwatch = { + dur: 11, still: 8.6, + build(svg, api) { + txt(svg, 44, 62, "Anthropic's speed claim", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Announcement of 7 October 2026", { "font-size": 20, "font-weight": 500, fill: MUTE }); + txt(svg, 756, 60, "tap the stopwatch", { "text-anchor": "end", "font-size": 20, fill: MUTE, "font-weight": 500 }); + const CX = 236, CY = 318, R = 138; + const w = el("g", {}, svg); + const crown = el("g", {}, w); + el("rect", { x: -20, y: -R - 44, width: 40, height: 30, rx: 8, fill: SOFT, stroke: INK, "stroke-width": 3 }, crown); + el("rect", { x: -11, y: -R - 18, width: 22, height: 20, fill: "#fff", stroke: INK, "stroke-width": 3 }, w); + el("rect", { x: -13, y: -R - 30, width: 26, height: 22, rx: 6, fill: LIL, stroke: INK, "stroke-width": 3, transform: "rotate(42)" }, w); + el("circle", { r: R, fill: YEL, stroke: INK, "stroke-width": 3 }, w); + el("circle", { r: R - 22, fill: "#fff", stroke: INK, "stroke-width": 3 }, w); + for (let i = 0; i < 60; i++) { const big = i % 5 === 0; el("path", { d: `M0 ${-(R - 30)} V${-(R - (big ? 42 : 37))}`, stroke: big ? INK : GREY, "stroke-width": big ? 3 : 2, "stroke-linecap": "round", transform: `rotate(${i * 6})` }, w); } + const arc = el("circle", { r: R - 54, fill: "none", stroke: SKY, "stroke-width": 10, "stroke-linecap": "round", pathLength: 360, transform: "rotate(-90)" }, w); + txt(w, 0, 50, "Haiku 5.5", { "text-anchor": "middle", "font-size": 24, "font-weight": 700 }); + const hand = el("g", {}, w); + el("path", { d: `M0 22 V${-(R - 44)}`, stroke: INK, "stroke-width": 5, "stroke-linecap": "round" }, hand); + el("circle", { r: 10, fill: BLUE, stroke: INK, "stroke-width": 3 }, w); + el("circle", { r: R + 10, fill: "#fff", "fill-opacity": 0, stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, w); + button(w, "Stopwatch for Haiku 5.5. Press to run it again.", () => api.seek(0.2)); + const label = (y, fill, a, b, rot) => { + const g = el("g", {}, svg); + el("rect", { x: 0, y: -52, width: 300, height: 104, rx: 16, fill, stroke: INK, "stroke-width": 3 }, g); + el("circle", { cx: 26, cy: 0, r: 8, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + txt(g, 52, -8, a, { "font-size": 24, "font-weight": 700 }); + txt(g, 52, 24, b, { "font-size": 24, "font-weight": 500 }); + return { g, y, rot }; + }; + const A = label(252, YEL, "fastest to date", "at standard speed", -2.5), B = label(384, LIL, "Opus in Fast Mode", "is quicker", 2); + const cap = strip(svg, 506, 640); + return (t) => { + const run = oc(seg(t, 0.8, 3.6)), deg = 312 * run, out = 1 - seg(t, 10.4, 10.9); + hand.setAttribute("transform", `rotate(${deg * out})`); + arc.setAttribute("stroke-dasharray", `${Math.max(0.01, deg * out)} 400`); arc.setAttribute("opacity", deg * out < 2 ? 0 : 1); + const press = Math.sin(seg(t, 0.55, 0.85) * Math.PI) + Math.sin(seg(t, 3.55, 3.85) * Math.PI); + crown.setAttribute("transform", `translate(0 ${8 * press})`); + const jolt = Math.sin(seg(t, 3.6, 4.0) * Math.PI) * 0.03; + tf(w, CX, CY, 1 + jolt); + [[A, 4.2], [B, 5.4]].forEach(([l, s]) => { const k = ob(seg(t, s, s + 0.6)) * out; tf(l.g, 820 - 384 * k, l.y, 1, l.rot * k); l.g.setAttribute("opacity", k <= 0 ? 0 : 1); }); + cap.textContent = t < 5.4 || t > 10.4 ? "Anthropic calls Haiku 5.5 its fastest at standard speed" : "Anthropic's claim. We have not tried it."; + }; + }, + }; + + /* ---- 2. a small measuring cup fills, spills over its lip and carries on into a larger one ---- */ + SC.cups = { + dur: 12.5, still: 9.6, + build(svg) { + txt(svg, 44, 62, "One model, two price tiers", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Anthropic's prices for Haiku 5.5, per 1 million tokens", { "font-size": 20, "font-weight": 500, fill: MUTE }); + const G = 404; + const s1 = el("path", { stroke: SKY, "stroke-width": 12, fill: "none", "stroke-linecap": "round" }, svg); + const s2 = el("path", { stroke: SKY, "stroke-width": 12, fill: "none", "stroke-linecap": "round" }, svg); + /* tap */ + el("path", { d: "M50 150 H186 Q210 150 210 172 V184", stroke: INK, "stroke-width": 28, fill: "none", "stroke-linejoin": "round" }, svg); + el("path", { d: "M50 150 H186 Q210 150 210 172 V184", stroke: SOFT, "stroke-width": 22, fill: "none", "stroke-linejoin": "round" }, svg); + el("rect", { x: 44, y: 126, width: 14, height: 48, rx: 5, fill: LIL, stroke: INK, "stroke-width": 3 }, svg); + el("path", { d: "M128 136 V124", stroke: INK, "stroke-width": 3 }, svg); + const wh = el("rect", { x: 104, y: 112, width: 48, height: 13, rx: 6.5, fill: YEL, stroke: INK, "stroke-width": 3 }, svg); + const cup = (x, top, b, w, lip) => { + const r = 16, hw = lip ? 30 : 46, hh = lip ? 24 : 36; + const liq = el("path", { fill: SKY }, svg); + const n = lip ? 3 : 5; + for (let i = 1; i <= n; i++) el("path", { d: `M${x + 3} ${b - (i * (b - top)) / (n + 1)} h${i % 2 ? 18 : 28}`, stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round" }, svg); + el("path", { d: `M${x + w} ${top + 26} q${hw} 0 ${hw} ${hh} q0 ${hh} -${hw} ${hh}`, stroke: INK, "stroke-width": 3, fill: "none", "stroke-linecap": "round", transform: lip ? `translate(${2 * x + w} 0) scale(-1 1)` : "" }, svg); + el("path", { d: `M${x} ${top} V${b - r} Q${x} ${b} ${x + r} ${b} H${x + w - r} Q${x + w} ${b} ${x + w} ${b - r} V${top}` + (lip ? ` l14 -8` : ""), stroke: INK, "stroke-width": 3, fill: "none", "stroke-linecap": "round", "stroke-linejoin": "round" }, svg); + return (lv) => { const y = b - (b - top) * lv; liq.setAttribute("d", lv <= 0.01 ? "" : `M${x + 1.5} ${y} H${x + w - 1.5} V${b - r} Q${x + w - 1.5} ${b - 1.5} ${x + w - r} ${b - 1.5} H${x + r} Q${x + 1.5} ${b - 1.5} ${x + 1.5} ${b - r} Z`); return y; }; + }; + el("rect", { x: 122, y: 292, width: 176, height: G - 292, rx: 12, fill: SOFT, stroke: INK, "stroke-width": 3 }, svg); + const small = cup(146, 200, 292, 128, true), big = cup(420, 262, G, 270, false); + el("path", { d: `M44 ${G} H756`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + const card = (cx, fill, head, a, b) => { + const g = el("g", {}, svg); + el("rect", { x: -150, y: 0, width: 300, height: 128, rx: 16, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + el("path", { d: "M-150 42 H150 V16 Q150 0 134 0 H-134 Q-150 0 -150 16 Z", fill, stroke: INK, "stroke-width": 3, "stroke-linejoin": "round" }, g); + txt(g, 0, 29, head, { "text-anchor": "middle", "font-size": 20, "font-weight": 700 }); + [["input", a, 76], ["output", b, 110]].forEach(([k, v, y]) => { txt(g, -126, y, k, { "font-size": 22, "font-weight": 500 }); txt(g, 126, y, v, { "text-anchor": "end", "font-family": MONO, "font-size": 26, "font-weight": 700 }); }); + return { g, cx }; + }; + const c1 = card(210, YEL, "prompts up to 100k tokens", "$0.10", "$0.50"), c2 = card(570, LIL, "prompts over 100k tokens", "$0.50", "$2.50"); + return (t) => { + const out = 1 - seg(t, 11.6, 12.3); + const on = seg(t, 0.5, 0.8) * (1 - seg(t, 8.2, 8.5)); + const f1 = seg(t, 0.8, 3.4), f2 = seg(t, 4.0, 8.4); + const y1 = small(0.93 * f1 * out), y2 = big(0.72 * f2 * out); + wh.setAttribute("transform", `rotate(${-9 * (seg(t, 0.3, 0.6) - seg(t, 8.0, 8.3))} 128 118)`); + const wob = Math.sin(t * 20) * 1.2; + s1.setAttribute("d", `M210 188 V${y1 + 4}`); s1.setAttribute("opacity", on); s1.setAttribute("stroke-width", 12 + wob); + const over = seg(t, 3.5, 3.9) * (1 - seg(t, 8.5, 8.8)); + s2.setAttribute("d", `M284 198 Q${284 + 204 * over} 184 ${284 + 216 * over} ${198 + (y2 - 194) * over}`); s2.setAttribute("opacity", over <= 0 ? 0 : 1); s2.setAttribute("stroke-width", 12 - wob); + [[c1, 3.4], [c2, 5.2]].forEach(([c, s]) => { const k = ob(seg(t, s, s + 0.5)) * out; tf(c.g, c.cx, 432 + 20 * (1 - k), 0.7 + 0.3 * k); c.g.setAttribute("opacity", cl(k * 3)); }); + }; + }, + }; + + /* ---- 3. four paper planes, one benchmark at a time; each lands as far as its score ---- */ + SC.planes = { + dur: 14, still: 4.8, + build(svg, api) { + const M = [["Haiku 5.5", YEL], ["Haiku 4.5", SKY], ["GPT-6 Luna", LIL], ["Sonnet 5.5", BLUE]]; + const B = [ + ["GDPval-AA v2.1", ["1620", "735", "1437", "1840"], 2000, ""], + ["OSWorld 2.1, offline subset", ["72.4%", "15.7%", "48.9%", "83.9%"], 100, ""], + ["Humanity's Last Exam, no tools", ["45.9%", "10.2%", null, "56.9%"], 100, ""], + ["Terminal-Bench 4.0", ["39.2%", "0.0%", "16.4%", "70.6%"], 100, ""], + ["FrontierCode 1.1 (Main)", ["46.4%", null, "42.4%", "52.1%"], 100, "Xhigh effort"], + ]; + const N = B.length, P = 2.8, X0 = 236, LEN = 400, Y = [164, 232, 300, 368]; + const at = (b, i) => { const v = B[b][1][i]; return v == null ? { x: X0, nr: 1 } : { x: X0 + (parseFloat(v) / B[b][2]) * LEN, nr: 0 }; }; + txt(svg, 44, 62, "Anthropic's own figures", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Five rows we selected. We did not run them. NR means not reported.", { "font-size": 20, "font-weight": 500, fill: MUTE }); + el("path", { d: `M${X0 - 40} 132 V412`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + el("rect", { x: X0 - 40, y: 412, width: LEN + 64, height: 26, rx: 8, fill: SOFT, stroke: INK, "stroke-width": 3 }, svg); + for (let i = 1; i < 20; i++) el("path", { d: `M${X0 + (i * LEN) / 20} 412 v${i % 2 ? 9 : 15}`, stroke: INK, "stroke-width": 2.5 }, svg); + const pl = M.map(([name, col], i) => { + txt(svg, 44, Y[i] + 8, name, { "font-size": 22, "font-weight": 700 }); + const trail = el("path", { stroke: GREY, "stroke-width": 3, fill: "none", "stroke-linecap": "round", "stroke-dasharray": "2 9" }, svg); + const g = el("g", {}, svg); + const body = el("path", { d: "M-34 -15 L30 0 L-34 15 L-22 0 Z", fill: col, stroke: INK, "stroke-width": 3, "stroke-linejoin": "round" }, g); + el("path", { d: "M-22 0 H30", stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round" }, g); + const v = txt(svg, 0, Y[i] + 9, "", { "font-size": 24, "font-weight": 700 }); + const note = txt(svg, 0, Y[i] + 8, "", { "font-size": 20, "font-weight": 500, fill: MUTE }); + return { trail, g, body, v, note, col }; + }); + el("rect", { x: 172, y: 478, width: 456, height: 52, rx: 26, fill: SOFT }, svg); + const name = txt(svg, 400, 512, "", { "text-anchor": "middle", "font-size": 22, "font-weight": 700 }); + let user = null, shown = 0; + const step = (d) => { user = { from: shown, to: (shown + d + N) % N, t0: api.now() }; api.poke(); }; + [[-1, 130, "Previous benchmark", "M5 -9 L-5 0 L5 9"], [1, 670, "Next benchmark", "M-5 -9 L5 0 L-5 9"]].forEach(([d, x, label, arrow]) => { + const g = el("g", { transform: `translate(${x} 504)` }, svg); + el("circle", { r: 32, fill: "none", stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, g); + el("circle", { r: 26, fill: "#fff", stroke: INK, "stroke-width": 3 }, g); + el("path", { d: arrow, stroke: INK, "stroke-width": 3.5, fill: "none", "stroke-linecap": "round", "stroke-linejoin": "round" }, g); + button(g, label, () => step(d)); + g.addEventListener("keydown", (e) => { const s = e.key === "ArrowRight" || e.key === "ArrowUp" ? 1 : e.key === "ArrowLeft" || e.key === "ArrowDown" ? -1 : 0; if (s) { e.preventDefault(); step(s); } }); + }); + return (t, now) => { + let idx, prev, u; + if (user) { idx = user.to; prev = user.from; u = cl((now - user.t0) / P); } + else { idx = Math.floor(t / P) % N; prev = (idx + N - 1) % N; u = (t % P) / P; } + shown = idx; name.textContent = B[idx][0]; + pl.forEach((p, i) => { + const a = at(prev, i), c = at(idx, i), k = io(seg(u, 0.02 + i * 0.03, 0.36 + i * 0.03)), x = a.x + (c.x - a.x) * k, lift = Math.sin(k * Math.PI); + tf(p.g, x, Y[i] - 16 * lift, 1, -9 * lift * Math.sign(c.x - a.x || 1) * Math.cos(k * Math.PI)); + p.trail.setAttribute("d", x - X0 < 34 ? "" : `M${X0 - 24} ${Y[i]} H${x - 46}`); + const s = B[idx][1][i], show = seg(u, 0.4 + i * 0.03, 0.5 + i * 0.03); + p.v.textContent = s == null ? "NR" : s; p.v.setAttribute("x", c.x + 44); p.v.setAttribute("opacity", show); + const nt = i === 3 ? B[idx][3] : ""; p.note.textContent = nt; if (nt) { p.note.setAttribute("x", c.x + 44 + p.v.getComputedTextLength() + 12); p.note.setAttribute("opacity", show); } + p.body.setAttribute("fill", c.nr && k > 0.5 ? "#fff" : p.col); p.body.setAttribute("stroke-dasharray", c.nr && k > 0.5 ? "6 6" : "none"); + }); + }; + }, + }; + + /* ---- 4. a postcard rack turns; the card at the front flips to show what the alias gives there ---- */ + SC.postcards = { + dur: 14.6, still: 12.6, + build(svg, api) { + const C = [[["Claude Platform", "on AWS"], "Haiku 4.5", SKY, "Claude Platform on AWS"], [["Amazon", "Bedrock"], "Haiku 4.5", LIL, "Amazon Bedrock"], [["Google Cloud's", "Agent Platform"], "Haiku 4.5", YEL, "Google Cloud's Agent Platform"], [["Microsoft", "Foundry"], "Haiku 4.5", BLUE, "Microsoft Foundry"], [["Anthropic", "API"], "Haiku 5.5", MINT, "the Anthropic API"]]; + const N = C.length, P = 2.6, R = 262, CY = 296, W = 230, H = 276; + txt(svg, 44, 62, "Where the haiku alias points", { "font-size": 32, "font-weight": 700 }); + txt(svg, 44, 98, "Claude Code model docs, read 8 Oct 2026", { "font-size": 20, "font-weight": 500, fill: MUTE }); + txt(svg, 756, 60, "tap to turn", { "text-anchor": "end", "font-size": 20, fill: MUTE, "font-weight": 500 }); + el("ellipse", { cx: 400, cy: CY, rx: R, ry: 22, fill: "none", stroke: GREY, "stroke-width": 3, "stroke-dasharray": "2 9", "stroke-linecap": "round" }, svg); + const layer = el("g", {}, svg); + el("rect", { x: 34, y: 116, width: 732, height: 374, rx: 26, fill: "#fff", "fill-opacity": 0, stroke: INK, "stroke-width": 2, "stroke-dasharray": "3 7", class: "mg-ring" }, layer); + const pole = el("g", {}, layer); + el("rect", { x: 330, y: 462, width: 140, height: 20, rx: 10, fill: SOFT, stroke: INK, "stroke-width": 3 }, pole); + el("path", { d: "M400 128 V462", stroke: INK, "stroke-width": 6, "stroke-linecap": "round" }, pole); + el("circle", { cx: 400, cy: 124, r: 11, fill: YEL, stroke: INK, "stroke-width": 3 }, pole); + const cards = C.map(([nm, model, col]) => { + const g = el("g", {}, layer), flip = el("g", {}, g); + const front = el("g", {}, flip), back = el("g", {}, flip); + el("rect", { x: -W / 2, y: -H / 2, width: W, height: H, rx: 18, fill: "#fff", stroke: INK, "stroke-width": 3 }, front); + const ft = el("g", {}, front); + el("rect", { x: W / 2 - 62, y: -H / 2 + 18, width: 44, height: 52, rx: 6, fill: col, stroke: INK, "stroke-width": 3 }, ft); + nm.forEach((s, i) => txt(ft, 0, -8 + i * 34, s, { "text-anchor": "middle", "font-size": 26, "font-weight": 700 })); + el("path", { d: `M${-W / 2 + 28} 74 H${W / 2 - 28} M${-W / 2 + 28} 98 H${W / 2 - 70}`, stroke: SOFT, "stroke-width": 6, "stroke-linecap": "round" }, ft); + el("rect", { x: -W / 2, y: -H / 2, width: W, height: H, rx: 18, fill: model === "Haiku 5.5" ? MINT : SOFT, stroke: INK, "stroke-width": 3 }, back); + const bt = el("g", {}, back); + const pg = el("g", { transform: "translate(0 -92)" }, bt); + el("rect", { x: -52, y: -22, width: 104, height: 44, rx: 12, fill: "#fff", stroke: INK, "stroke-width": 3 }, pg); + txt(pg, 0, 8, "haiku", { "text-anchor": "middle", "font-family": MONO, "font-size": 24, "font-weight": 700 }); + txt(bt, 0, -34, "resolves to", { "text-anchor": "middle", "font-size": 24, "font-weight": 500 }); + txt(bt, 0, 14, model, { "text-anchor": "middle", "font-size": 40, "font-weight": 700 }); + el("path", { d: `M${-W / 2 + 28} 40 H${W / 2 - 28}`, stroke: INK, "stroke-width": 2.5, "stroke-linecap": "round", "stroke-dasharray": "2 8" }, bt); + nm.forEach((s, i) => txt(bt, 0, 78 + i * 30, s, { "text-anchor": "middle", "font-size": 24, "font-weight": 600 })); + return { g, flip, front, back, ft, bt }; + }); + const cap = strip(svg, 506, 700); + let last = 0, key = ""; + button(layer, "Postcard rack. Press to turn to the next provider.", () => { const k = Math.min(N - 1, Math.floor(last / P)); api.seek(((k + 1) % N) * P + 0.02); }); + return (t) => { + last = t; + const k = Math.min(N - 1, Math.floor(t / P)), u = t - k * P, pos = k - 1 + io(seg(u, 0, 0.9)), back = seg(t, N * P + 0.9, N * P + 1.4); + const order = [[0, pole]]; + cards.forEach((c, i) => { + const a = ((i - pos) * 2 * Math.PI) / N, z = Math.cos(a), s = 0.64 + 0.36 * (z + 1) / 2; + tf(c.g, 400 + R * Math.sin(a), CY + 16 * z, s); + const f = (i < k ? 1 : i === k ? io(seg(u, 1.05, 1.6)) : 0) * (1 - io(back)), sx = Math.abs(Math.cos(f * Math.PI)); + c.flip.setAttribute("transform", `scale(${Math.max(0.02, sx)} 1)`); + c.front.setAttribute("display", f > 0.5 ? "none" : ""); c.back.setAttribute("display", f > 0.5 ? "" : "none"); + const vis = seg(z, -0.05, 0.28); c.ft.setAttribute("opacity", vis); c.bt.setAttribute("opacity", vis); + order.push([z, c.g]); + }); + order.sort((p, q) => p[0] - q[0]); + const nk = order.map((o) => (o[1] === pole ? "p" : cards.findIndex((c) => c.g === o[1]))).join(""); + if (nk !== key) { key = nk; order.forEach((o) => layer.appendChild(o[1])); } + cap.textContent = u < 1.35 || back > 0 ? "In Claude Code, the haiku alias depends on your provider" : `On ${C[k][3]}, haiku resolves to ${C[k][1]}`; + }; + }, + }; + + /* ---- title card, 16 by 9: a row of dominoes part way through falling ---- */ + SC.hero = { + w: 960, h: 540, dur: 9, still: 2.7, + build(svg) { + const kg = el("g", { transform: "translate(64 62)" }, svg); + const kr = el("rect", { x: 0, y: 0, height: 40, rx: 20, fill: SOFT, stroke: INK, "stroke-width": 2.5 }, kg); + const kt = txt(kg, 20, 28, "Concepts", { "font-size": 22, "font-weight": 700 }); + kr.setAttribute("width", kt.getComputedTextLength() + 40); + const title = txt(svg, 62, 206, "Claude Haiku 5.5", { "font-size": 96, "font-weight": 700 }); + el("path", { d: `M66 236 H${62 + title.getComputedTextLength() - 6}`, stroke: YEL, "stroke-width": 14, "stroke-linecap": "round" }, svg); + txt(svg, 64, 292, "Checked 8 Oct 2026", { "font-size": 26, "font-weight": 500, fill: MUTE }); + const G = 478, DW = 20, DH = 96, D = 52, n = 15, TH = (Math.acos(DW / D) * 180) / Math.PI; + el("path", { d: `M64 ${G} H896`, stroke: INK, "stroke-width": 3, "stroke-linecap": "round" }, svg); + const COL = [YEL, SKY, LIL, MINT, BLUE]; + const ds = Array.from({ length: n }, (_, i) => { + const g = el("g", {}, svg); + el("rect", { x: -DW, y: -DH, width: DW, height: DH, rx: 5, fill: COL[i % 5], stroke: INK, "stroke-width": 3 }, g); + return g; + }); + const RAD = Math.PI / 180; + const fits = (a, b) => { + const px = DH * Math.sin(a), py = DH * Math.cos(a), bx = D - DW * Math.cos(b), by = DW * Math.sin(b); + const along = (px - bx) * Math.sin(b) + (py - by) * Math.cos(b), side = (px - bx) * Math.cos(b) - (py - by) * Math.sin(b); + if (along <= DH) return side <= 0.3; + return (bx + DH * Math.sin(b)) * Math.cos(a) - (by + DH * Math.cos(b)) * Math.sin(a) >= -0.3; + }; + return (t) => { + const up = 1 - io(seg(t, 7.6, 8.6)); + let next = null; + for (let i = n - 1; i >= 0; i--) { + const k = seg(t, 0.6 + i * 0.24, 1.3 + i * 0.24); + let a = 90 * k * k * RAD; + if (next != null && !fits(a, next)) { let lo = 0, hi = a; for (let j = 0; j < 22; j++) { const m = (lo + hi) / 2; if (fits(m, next)) lo = m; else hi = m; } a = lo; } + next = a; + ds[i].setAttribute("transform", `translate(${104 + i * D} ${G}) rotate(${(a / RAD) * up})`); + } + }; + }, + }; + + /* ---- player ---- */ + const q = new URLSearchParams(location.search), frozen = q.has("mgstill"), fixed = parseFloat(q.get("mgstill")); + const calm = matchMedia("(prefers-reduced-motion: reduce)").matches; + const mount = (fig) => { + const sc = SC[fig.dataset.scene], img = fig.querySelector("img"); if (!sc || !img) return; + const W = sc.w || 800, H = sc.h || 600; + const svg = el("svg", { viewBox: `0 0 ${W} ${H}`, class: "mg-svg", role: "group", "aria-label": img.alt || "" }); + svg.style.aspectRatio = `${W} / ${H}`; + el("rect", { width: W, height: H, fill: "#fff" }, svg); + img.replaceWith(svg); + let off = 0, raf = 0, seen = false, held = calm; + const base = performance.now(), clock = () => (performance.now() - base) / 1000; + let draw = () => {}; + const frame = () => { const now = clock(); draw(held ? sc.still : (((now + off) % sc.dur) + sc.dur) % sc.dur, now); }; + const tick = () => { cancelAnimationFrame(raf); frame(); if (seen && !held) raf = requestAnimationFrame(tick); }; + const api = { now: clock, seek: (s) => { off = s - clock(); held = false; tick(); }, poke: () => { held = false; tick(); } }; + draw = sc.build(svg, api); + if (frozen) { draw(isNaN(fixed) ? sc.still : fixed, 0); return; } + new IntersectionObserver(([e]) => { + seen = e.isIntersecting; + if (seen) { tick(); if (!fig.dataset.sent && window.posthog) { fig.dataset.sent = 1; window.posthog.capture("blog_scene_view", { slug: location.pathname.split("/").filter(Boolean).pop(), scene: fig.dataset.scene }); } } + else cancelAnimationFrame(raf); + }, { threshold: 0.25 }).observe(fig); + svg.addEventListener("click", () => { if (!fig.dataset.hit && window.posthog) { fig.dataset.hit = 1; window.posthog.capture("blog_scene_interact", { slug: location.pathname.split("/").filter(Boolean).pop(), scene: fig.dataset.scene }); } }); + frame(); + }; + const go = () => document.querySelectorAll("figure.mg[data-scene]").forEach(mount); + const fonts = document.fonts && document.fonts.load ? Promise.all([document.fonts.load('700 32px "Space Grotesk"'), document.fonts.load('500 20px "Space Grotesk"'), document.fonts.load('700 20px "JetBrains Mono"')]).catch(() => {}) : Promise.resolve(); + const start = () => fonts.then(go, go); + if (document.readyState === "loading") document.addEventListener("DOMContentLoaded", start); else start(); +})(); diff --git a/docs/blog/assets/media/claude-haiku-5-5/planes.png b/docs/blog/assets/media/claude-haiku-5-5/planes.png new file mode 100644 index 000000000..b57c59d02 Binary files /dev/null and b/docs/blog/assets/media/claude-haiku-5-5/planes.png differ diff --git a/docs/blog/assets/media/claude-haiku-5-5/postcards.png b/docs/blog/assets/media/claude-haiku-5-5/postcards.png new file mode 100644 index 000000000..24b9bcac6 Binary files /dev/null and b/docs/blog/assets/media/claude-haiku-5-5/postcards.png differ diff --git a/docs/blog/assets/media/claude-haiku-5-5/stopwatch.png b/docs/blog/assets/media/claude-haiku-5-5/stopwatch.png new file mode 100644 index 000000000..933b39827 Binary files /dev/null and b/docs/blog/assets/media/claude-haiku-5-5/stopwatch.png differ diff --git a/docs/blog/can-claude-code-agents-talk-to-each-other/index.html b/docs/blog/can-claude-code-agents-talk-to-each-other/index.html index 305449b3c..ad6709d19 100644 --- a/docs/blog/can-claude-code-agents-talk-to-each-other/index.html +++ b/docs/blog/can-claude-code-agents-talk-to-each-other/index.html @@ -341,50 +341,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/claude-code-subagents-vs-multi-agent-harness/index.html b/docs/blog/claude-code-subagents-vs-multi-agent-harness/index.html index 3d5b96cf9..17b8741c9 100644 --- a/docs/blog/claude-code-subagents-vs-multi-agent-harness/index.html +++ b/docs/blog/claude-code-subagents-vs-multi-agent-harness/index.html @@ -217,7 +217,7 @@

How does Cl

It reads the description field. When your request matches a description, Claude hands off the task with a delegation prompt it writes itself. Put “use proactively” in the description if you want it to delegate without being asked.

You can also force it: name the subagent in plain words, @-mention it with @agent-test-runner, or run a whole session as that agent with claude --agent test-runner. The subagent starts fresh: it sees its own prompt, your CLAUDE.md files and the delegation message, not your conversation history.

How do you set a subagent’s model and tools?

-

Use model and tools in the front matter. The model is resolved in this order: a model Claude passes when it spawns the subagent, then the model field, then the CLAUDE_CODE_SUBAGENT_MODEL variable, then your session’s model. Since 2.1.251 that variable is only a default. To force one model on every subagent, also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 (2.1.257 and later). Run /tasks to see which model each subagent is on.

+

Use model and tools in the front matter. The model is resolved in this order: a model Claude passes when it spawns the subagent, then the model field, then the CLAUDE_CODE_SUBAGENT_MODEL variable, then your session’s model. Since 2.1.251 that variable is only a default. To force one model on every subagent, also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 (2.1.257 and later). Run /tasks to see which model each subagent is on. The newest Haiku is Claude Haiku 5.5.

tools is an allowlist and disallowedTools a denylist. A disallowedTools entry such as Bash(git push *) removes the whole Bash tool, so block single commands with a deny rule in settings instead. The Task tool was renamed Agent in 2.1.63, and old Task(...) rules still work.

Do Claude Code subagents run in the background?

Yes, by default in an interactive session. Background became the default in 2.1.198, so the subagents Claude spawns run while you keep typing. Since 2.1.232 (13 Aug 2026) fork mode is on, so Claude can no longer ask for the foreground. A background subagent gets a smaller built-in tool set, and its permission prompts appear in your main session with its name on them. /subtask starts a fork, which is a subagent that inherits your whole conversation. Up to 20 subagents can run at once; CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS changes that.

diff --git a/docs/blog/claude-haiku-5-5/index.html b/docs/blog/claude-haiku-5-5/index.html new file mode 100644 index 000000000..61ff10a15 --- /dev/null +++ b/docs/blog/claude-haiku-5-5/index.html @@ -0,0 +1,492 @@ + + + + + + + + +Claude Haiku 5.5: price, benchmarks and where it fits — Munder Difflin Blog + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ +
+
+ +

Claude Haiku 5.5: price, benchmarks and where it fits

+

Claude Haiku 5.5 is Anthropic's new small model. Its tiered price, Anthropic's own scores and how to pick it in Claude Code, checked 8 Oct 2026.

+
+ + Chaitanya Giri + + + + 5 min read + Concepts +
+
+ +
Title card on white. Claude Haiku 5.5 in large ink type over a yellow underline, a Concepts kicker, the line Checked 8 Oct 2026, and a row of coloured dominoes part way through falling
+ +
+
+

Claude Haiku 5.5 is Anthropic’s new small model: much cheaper than Haiku 4.5 on prompts up to 100,000 tokens, and behind Sonnet 5.5 on every row of Anthropic’s own benchmark table. Anthropic released it on 7 October 2026. We have not tried it. Checked 8 Oct 2026.

+

Munder Difflin is free and open source: a desktop app that runs a team of coding agents such as Claude Code, Codex and Gemini CLI on your own computer. A cheaper small model matters when several agents run at once, which is why we cover it. The install guide covers setup and the Concepts hub explains the terms.

+

What is Claude Haiku 5.5?

+

It is Anthropic’s new small model, with the model ID claude-haiku-5-5. Anthropic’s announcement, dated October 7, 2026, calls it “the cheapest, fastest, and most capable small model we’ve ever released” and says it is designed for high-volume, cost-sensitive tasks.

+

The speed claim has a footnote. Anthropic says Haiku 5.5 is its fastest model to date at each model’s standard speed, although it runs less quickly than its Opus models in Fast Mode.

+

Anthropic’s model page, read on 8 Oct 2026, lists a 1M token context window and a maximum output of 128K tokens. The announcement adds that this is the first Haiku-class model with an adjustable effort setting.

+

The Hacker News thread had more than 300 points and 140 comments when we checked on 8 Oct 2026.

+
Animation. A stopwatch labelled Haiku 5.5 starts, its hand sweeps round and stops. Two paper labels slide in beside it: one reads fastest to date at standard speed, the other reads Opus in Fast Mode is quicker.
Anthropic calls Haiku 5.5 its fastest model to date at standard speed, with Opus in Fast Mode quicker. Announcement of 7 October 2026.
+

What does Claude Haiku 5.5 cost?

+

It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, as of 8 Oct 2026, and five times that above. These are the prices in the table in Anthropic’s announcement, per 1 million tokens.

+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Price per 1M tokensHaiku 5.5, prompts up to 100kHaiku 5.5, prompts over 100kHaiku 4.5Sonnet 5.5
Input$0.10$0.50$1.00$2.00
Output$0.50$2.50$5.00$10.00
Cache writes$0.125$0.625$1.25$2.50
Cache reads$0.01$0.05$0.10$0.10
+
+

Anthropic says prompts up to 100,000 tokens made up around 90% of requests to its previous Haiku model. Its headline is that Haiku 5.5 costs around 75% less to run on average. A footnote explains the gap: the price is 90% lower for requests up to 100,000 tokens, less for larger ones, and an updated tokenizer uses slightly more tokens per task. The model page says the same text counts as approximately 30% more tokens than on Haiku 4.5.

+

The model page also lists a 50% Batch API discount on input and output. For the wider bill, see how much Claude Code costs.

+
Animation. A tap fills a small measuring cup, and a card under it headed prompts up to 100k tokens shows $0.10 input and $0.50 output. The water spills over the lip into a larger cup, and its card headed prompts over 100k tokens shows $0.50 input and $2.50 output.
Anthropic's prices per 1 million tokens, in two tiers by prompt size. Checked 8 Oct 2026.
+

How does Claude Haiku 5.5 score on benchmarks?

+

On Anthropic’s own figures it is far ahead of Haiku 4.5 and behind Sonnet 5.5. The five rows below are our selection from the eight rows in the announcement table. We did not run them. NR means not reported.

+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Benchmark (as named by Anthropic)Haiku 5.5Haiku 4.5GPT-6 LunaSonnet 5.5
GDPval-AA v2.1162073514371840
OSWorld 2.1, offline subset72.4%15.7%48.9%83.9%
Humanity’s Last Exam, no tools45.9%10.2%NR56.9%
Terminal-Bench 4.039.2%0.0%16.4%70.6%
FrontierCode 1.1 (Main)46.4%NR42.4%52.1%
+
+

Sonnet 5.5 leads on all eight rows of the full table. Anthropic marks its FrontierCode figure for Sonnet 5.5 as Xhigh effort. Haiku 5.5 is ahead of GPT-6 Luna on all six rows where Anthropic reports both.

+

The coding gap is the one to notice. On Terminal-Bench 4.0, Haiku 5.5 is at 39.2% against 70.6% for Sonnet 5.5. Anthropic says so itself: “Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0.”

+
Animation. Four paper planes for Haiku 5.5, Haiku 4.5, GPT-6 Luna and Sonnet 5.5 glide along a ruled strip and land at distances set by their scores on one benchmark at a time. Arrow buttons step through the five benchmarks.
Anthropic's own figures, five rows we selected from its table of 7 October 2026. We did not run them.
+

Where can you use Claude Haiku 5.5?

+

Anthropic says it “is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure”. The model page lists the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. The ID is claude-haiku-5-5, except on Amazon Bedrock, where it is anthropic.claude-haiku-5-5.

+

Neither page says which Claude app plans include it, so we leave that out.

+

How do you use Claude Haiku 5.5 in Claude Code?

+

Run /model claude-haiku-5-5 in a session, or start with claude --model claude-haiku-5-5. That wording is from Claude Code’s model configuration docs, read on 8 Oct 2026. The same page says to use v2.1.293 or later with Haiku 5.5.

+

Three details from the docs:

+
    +
  • The alias depends on your provider. haiku resolves to Haiku 5.5 on the Anthropic API. On Claude Platform on AWS, Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry it resolves to Haiku 4.5.
    • +
  • Long prompts cost more. The docs repeat that a Haiku 5.5 request costs more per token when its prompt is longer than 100K tokens.
    • +
  • Subagents take a model. The subagent docs say the model field accepts an alias such as haiku, a full model ID or inherit. CLAUDE_CODE_SUBAGENT_MODEL sets a default.
    • +
+

Our post on Claude Code subagents walks through that file.

+
Animation. A postcard rack turns slowly. Each card names a provider, and as it faces front it flips to show which model the haiku alias gives: Haiku 5.5 on the Anthropic API card, Haiku 4.5 on the other cards.
In Claude Code the haiku alias resolves to Haiku 5.5 on the Anthropic API and to Haiku 4.5 on the other listed providers. Docs read 8 Oct 2026.
+

What does it mean if you run a team of coding agents?

+

We think it makes the split between a planner and its helpers cheaper to try. Anthropic says Haiku 5.5 “pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work”, and that it suits narrowly scoped tasks like compaction, summarization or subagent work.

+

Two customer quotes from the announcement, the first on subagents and the second on speed. Rogo: “The short and high-volume work is where Claude Haiku 5.5 fits for us, like quick lookups, subagents, and summaries.” Asana: “Compared with the model we use today, we saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn.”

+

Our view: put a bigger model on planning and hard debugging, and a small fast one on searching, summarising and running tests. The small one fetches, the big one decides. This is an idea, not a test. We have not tried Haiku 5.5. Our piece on model routing explains the reasoning.

+

What should you do now?

+

Our view: try it on routine work first, and keep your bigger model where the task is hard.

+
    +
  1. Update Claude Code to v2.1.293 or later.
    2. +
  2. Set model: haiku on one subagent that only searches or summarises.
    3. +
  3. Watch prompt size. The lower price applies up to 100,000 tokens.
    4. +
  4. Compare results on your own repo before trusting any table.
    5. +
+ + + + +
+

FAQ

+ +
+

How much does Claude Haiku 5.5 cost?

+

Anthropic's announcement of 7 October 2026 lists $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 input and $2.50 output for prompts over 100,000 tokens. Cache reads are $0.01 and $0.05. Checked 8 Oct 2026.

+
+ +
+

Is Claude Haiku 5.5 better than Haiku 4.5?

+

On Anthropic's own figures, yes on every row where both have a score. For example, the announcement lists 72.4% for Haiku 5.5 and 15.7% for Haiku 4.5 on the offline subset of OSWorld 2.1. We did not run these tests.

+
+ +
+

Is Claude Haiku 5.5 better than Sonnet 5.5?

+

No. Sonnet 5.5 is ahead on every row of Anthropic's own benchmark table, and the announcement says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks. Haiku 5.5 is the cheaper and faster option.

+
+ +
+

What is the model ID for Claude Haiku 5.5?

+

It is claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock, according to Anthropic's model page read on 8 Oct 2026.

+
+ +
+

How do I use Claude Haiku 5.5 in Claude Code?

+

Claude Code's model docs say to run /model claude-haiku-5-5 in a session, or start with claude --model claude-haiku-5-5. The same page says to use Claude Code v2.1.293 or later with Haiku 5.5. Read on 8 Oct 2026.

+
+ +
+ + + +
+ + +
+ + + + +
+ + +
+ +
+ + + + + + diff --git a/docs/blog/do-more-with-less-model-routing/index.html b/docs/blog/do-more-with-less-model-routing/index.html index 123e47112..873025145 100644 --- a/docs/blog/do-more-with-less-model-routing/index.html +++ b/docs/blog/do-more-with-less-model-routing/index.html @@ -330,50 +330,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/feed.xml b/docs/blog/feed.xml index 2d9d0f0de..4550309b3 100644 --- a/docs/blog/feed.xml +++ b/docs/blog/feed.xml @@ -4,9 +4,17 @@ Guides, deep dives, and comparisons on running multi-agent Claude Code: orchestration, agent memory, automation, and the tooling landscape. - 2026-10-07T00:00:00.000Z + 2026-10-08T00:00:00.000Z https://munderdiffl.in/blog/ Chaitanya Girihttps://munderdiffl.in + + Claude Haiku 5.5: price, benchmarks and where it fits + + https://munderdiffl.in/blog/claude-haiku-5-5/ + 2026-10-08T00:00:00.000Z + 2026-10-08T00:00:00.000Z + Claude Haiku 5.5 is Anthropic's new small model. Its tiered price, Anthropic's own scores and how to pick it in Claude Code, checked 8 Oct 2026. + Agent Tools Today, 7 Oct 2026: Mistral Large 4, EmbeddingGemma 2, OpenAI Decisions API @@ -479,12 +487,4 @@ 2026-09-10T00:00:00.000Z Munder Difflin 0.5.2 is out: how to install it, how to start Pro, what Pro adds, what changed since 0.4.6, and the end of signed Community builds. - - Launching Munder Difflin v0.4.6: It Speaks Chinese and Arabic Now - - https://munderdiffl.in/blog/launching-munder-difflin-v0-4-6/ - 2026-08-27T00:00:00.000Z - 2026-08-27T00:00:00.000Z - v0.4.6 adds Chinese and Arabic interfaces with right-to-left support, ships its fonts inside the app so nothing is fetched from Google, and makes the updater actually update itself instead of pointing at a download page. - diff --git a/docs/blog/generative-engine-optimization-techniques/index.html b/docs/blog/generative-engine-optimization-techniques/index.html index 4e4662117..327adaf05 100644 --- a/docs/blog/generative-engine-optimization-techniques/index.html +++ b/docs/blog/generative-engine-optimization-techniques/index.html @@ -335,50 +335,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/how-ai-agents-verify-their-own-work/index.html b/docs/blog/how-ai-agents-verify-their-own-work/index.html index 275de503d..af89e5c00 100644 --- a/docs/blog/how-ai-agents-verify-their-own-work/index.html +++ b/docs/blog/how-ai-agents-verify-their-own-work/index.html @@ -329,50 +329,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/how-much-does-claude-code-cost/index.html b/docs/blog/how-much-does-claude-code-cost/index.html index a972caac3..9482aad43 100644 --- a/docs/blog/how-much-does-claude-code-cost/index.html +++ b/docs/blog/how-much-does-claude-code-cost/index.html @@ -416,50 +416,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/index.html b/docs/blog/index.html index 764991cd3..03cbbc92b 100644 --- a/docs/blog/index.html +++ b/docs/blog/index.html @@ -135,13 +135,13 @@

The Munder Difflin Blog.

building a self-coordinating hive of agents with a GOD orchestrator you talk to.

-
+
+ + + +
-
- + -
+
+
+ - -
- - - -
-
-
+
diff --git a/docs/blog/is-claude-code-max-worth-it/index.html b/docs/blog/is-claude-code-max-worth-it/index.html index bc548552d..e182a5567 100644 --- a/docs/blog/is-claude-code-max-worth-it/index.html +++ b/docs/blog/is-claude-code-max-worth-it/index.html @@ -411,50 +411,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/is-claude-code-open-source/index.html b/docs/blog/is-claude-code-open-source/index.html index 8af7e4df3..53930f3c1 100644 --- a/docs/blog/is-claude-code-open-source/index.html +++ b/docs/blog/is-claude-code-open-source/index.html @@ -292,50 +292,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/is-ollama-good-for-coding/index.html b/docs/blog/is-ollama-good-for-coding/index.html index 2fc17fe90..bb7ced766 100644 --- a/docs/blog/is-ollama-good-for-coding/index.html +++ b/docs/blog/is-ollama-good-for-coding/index.html @@ -332,50 +332,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/long-running-agents-the-2026-shift/index.html b/docs/blog/long-running-agents-the-2026-shift/index.html index bd41cd7fa..1ac0c53d3 100644 --- a/docs/blog/long-running-agents-the-2026-shift/index.html +++ b/docs/blog/long-running-agents-the-2026-shift/index.html @@ -350,50 +350,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/loop-engineering-for-ai-agents/index.html b/docs/blog/loop-engineering-for-ai-agents/index.html index 7685826dd..a97d15bf6 100644 --- a/docs/blog/loop-engineering-for-ai-agents/index.html +++ b/docs/blog/loop-engineering-for-ai-agents/index.html @@ -291,50 +291,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/mistral-large-4/index.html b/docs/blog/mistral-large-4/index.html index e90db9f12..50201f19d 100644 --- a/docs/blog/mistral-large-4/index.html +++ b/docs/blog/mistral-large-4/index.html @@ -346,50 +346,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/munder-difflin-faq/index.html b/docs/blog/munder-difflin-faq/index.html index ff20f9222..4e2041d7c 100644 --- a/docs/blog/munder-difflin-faq/index.html +++ b/docs/blog/munder-difflin-faq/index.html @@ -320,50 +320,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/open-source-ai-tools-on-purpose/index.html b/docs/blog/open-source-ai-tools-on-purpose/index.html index 347b6a803..adc89377f 100644 --- a/docs/blog/open-source-ai-tools-on-purpose/index.html +++ b/docs/blog/open-source-ai-tools-on-purpose/index.html @@ -322,50 +322,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/openai-decisions-api/index.html b/docs/blog/openai-decisions-api/index.html index 5212a4678..4b258a1df 100644 --- a/docs/blog/openai-decisions-api/index.html +++ b/docs/blog/openai-decisions-api/index.html @@ -361,50 +361,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/openai-text-watermarking/index.html b/docs/blog/openai-text-watermarking/index.html index 1fd073bca..44b334e22 100644 --- a/docs/blog/openai-text-watermarking/index.html +++ b/docs/blog/openai-text-watermarking/index.html @@ -372,50 +372,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/reflection-beam/index.html b/docs/blog/reflection-beam/index.html index 952ef0384..e5207bbc4 100644 --- a/docs/blog/reflection-beam/index.html +++ b/docs/blog/reflection-beam/index.html @@ -351,50 +351,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/review-agent-work-where-it-happens/index.html b/docs/blog/review-agent-work-where-it-happens/index.html index 2d6db30bd..b032083cf 100644 --- a/docs/blog/review-agent-work-where-it-happens/index.html +++ b/docs/blog/review-agent-work-where-it-happens/index.html @@ -282,50 +282,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/shareable-agent-roles/index.html b/docs/blog/shareable-agent-roles/index.html index 5427ee71f..5e0df8f3b 100644 --- a/docs/blog/shareable-agent-roles/index.html +++ b/docs/blog/shareable-agent-roles/index.html @@ -324,50 +324,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/tags/ai-agents/index.html b/docs/blog/tags/ai-agents/index.html index 3ed42cd1d..8bf8dc377 100644 --- a/docs/blog/tags/ai-agents/index.html +++ b/docs/blog/tags/ai-agents/index.html @@ -105,11 +105,27 @@

Tag

#AI Agents

-

17 posts tagged AI Agents.

+

18 posts tagged AI Agents.

+
diff --git a/docs/blog/tags/claude-code/index.html b/docs/blog/tags/claude-code/index.html index e8bceb701..f88bebb36 100644 --- a/docs/blog/tags/claude-code/index.html +++ b/docs/blog/tags/claude-code/index.html @@ -105,11 +105,27 @@

Tag

#Claude Code

-

73 posts tagged Claude Code.

+

74 posts tagged Claude Code.

-
+ +
diff --git a/docs/blog/tags/concepts/index.html b/docs/blog/tags/concepts/index.html index c232bacb6..acbeba930 100644 --- a/docs/blog/tags/concepts/index.html +++ b/docs/blog/tags/concepts/index.html @@ -105,11 +105,27 @@

Tag

#Concepts

-

35 posts tagged Concepts.

+

36 posts tagged Concepts.

+
diff --git a/docs/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/index.html b/docs/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/index.html index 28c03876a..0d5da76f2 100644 --- a/docs/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/index.html +++ b/docs/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/index.html @@ -358,50 +358,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/topics/comparisons/index.html b/docs/blog/topics/comparisons/index.html index 29be68fef..9023adfa0 100644 --- a/docs/blog/topics/comparisons/index.html +++ b/docs/blog/topics/comparisons/index.html @@ -116,9 +116,9 @@

comparisons

Honest side-by-sides with the other ways to guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/concepts/index.html b/docs/blog/topics/concepts/index.html index f1a46a27a..31f6cf1da 100644 --- a/docs/blog/topics/concepts/index.html +++ b/docs/blog/topics/concepts/index.html @@ -116,9 +116,9 @@

concepts

The ideas underneath the product — harnesses, guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 @@ -138,6 +138,22 @@

concepts

The ideas underneath the product — harnesses,

+
diff --git a/docs/blog/topics/guides/index.html b/docs/blog/topics/guides/index.html index aaf6db8d4..715760a32 100644 --- a/docs/blog/topics/guides/index.html +++ b/docs/blog/topics/guides/index.html @@ -116,9 +116,9 @@

guides

Step-by-step walkthroughs you can follow today guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/index.html b/docs/blog/topics/index.html index 10414a8bc..815f4be7a 100644 --- a/docs/blog/topics/index.html +++ b/docs/blog/topics/index.html @@ -122,20 +122,20 @@

guides

- - + +
-

comparisons

+

concepts

39 posts.

View topic →
 - - + +
-

concepts

-

38 posts.

+

comparisons

+

39 posts.

View topic →
 @@ -202,15 +202,15 @@

news

Multi-Agent88 - Claude Code73 + Claude Code74 Open Source53 Comparisons39 - Guides36 + Concepts36 - Concepts35 + Guides36 Internals26 @@ -226,7 +226,7 @@

news

CLI Agents18 - AI Agents17 + AI Agents18 Release16 diff --git a/docs/blog/topics/internals/index.html b/docs/blog/topics/internals/index.html index 861eeafb0..6592a3ca1 100644 --- a/docs/blog/topics/internals/index.html +++ b/docs/blog/topics/internals/index.html @@ -116,9 +116,9 @@

internals

How the machine actually works: postmortems, a guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/memory/index.html b/docs/blog/topics/memory/index.html index bc34f782d..e1183f39d 100644 --- a/docs/blog/topics/memory/index.html +++ b/docs/blog/topics/memory/index.html @@ -116,9 +116,9 @@

memory

How agents remember across sessions — and how a guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/news/index.html b/docs/blog/topics/news/index.html index b890e5fe8..683425f44 100644 --- a/docs/blog/topics/news/index.html +++ b/docs/blog/topics/news/index.html @@ -116,9 +116,9 @@

news

Agent Tools Today: a daily brief of the top five la guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/orchestration/index.html b/docs/blog/topics/orchestration/index.html index d48bf196f..7a847098d 100644 --- a/docs/blog/topics/orchestration/index.html +++ b/docs/blog/topics/orchestration/index.html @@ -116,9 +116,9 @@

orchestration

Patterns for making one boss agent run a f guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/story/index.html b/docs/blog/topics/story/index.html index ddb91485c..4c9bfc46b 100644 --- a/docs/blog/topics/story/index.html +++ b/docs/blog/topics/story/index.html @@ -116,9 +116,9 @@

story

Launches, releases, and the occasional confession guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/topics/use-cases/index.html b/docs/blog/topics/use-cases/index.html index 0ebe0e814..bdc44399e 100644 --- a/docs/blog/topics/use-cases/index.html +++ b/docs/blog/topics/use-cases/index.html @@ -116,9 +116,9 @@

use-cases

Real workflows the hive actually runs — writ guides47 - comparisons39 + concepts39 - concepts38 + comparisons39 internals19 diff --git a/docs/blog/voice-as-a-control-plane-for-agent-fleets/index.html b/docs/blog/voice-as-a-control-plane-for-agent-fleets/index.html index dd1094dd9..22e854074 100644 --- a/docs/blog/voice-as-a-control-plane-for-agent-fleets/index.html +++ b/docs/blog/voice-as-a-control-plane-for-agent-fleets/index.html @@ -284,50 +284,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-are-claude-code-agents/index.html b/docs/blog/what-are-claude-code-agents/index.html index 2f0cfc462..379847cb3 100644 --- a/docs/blog/what-are-claude-code-agents/index.html +++ b/docs/blog/what-are-claude-code-agents/index.html @@ -302,50 +302,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-a-multi-agent-harness/index.html b/docs/blog/what-is-a-multi-agent-harness/index.html index 360b2aacb..93b42f5a0 100644 --- a/docs/blog/what-is-a-multi-agent-harness/index.html +++ b/docs/blog/what-is-a-multi-agent-harness/index.html @@ -373,50 +373,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-aeo-for-dev-tools/index.html b/docs/blog/what-is-aeo-for-dev-tools/index.html index 13027fa8a..d95e7fab3 100644 --- a/docs/blog/what-is-aeo-for-dev-tools/index.html +++ b/docs/blog/what-is-aeo-for-dev-tools/index.html @@ -424,50 +424,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-an-agent-development-environment/index.html b/docs/blog/what-is-an-agent-development-environment/index.html index c22a08634..c6c145517 100644 --- a/docs/blog/what-is-an-agent-development-environment/index.html +++ b/docs/blog/what-is-an-agent-development-environment/index.html @@ -287,50 +287,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-an-mcp-server/index.html b/docs/blog/what-is-an-mcp-server/index.html index f9b08f301..ed1a13028 100644 --- a/docs/blog/what-is-an-mcp-server/index.html +++ b/docs/blog/what-is-an-mcp-server/index.html @@ -292,50 +292,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-chatgpt-dots/index.html b/docs/blog/what-is-chatgpt-dots/index.html index 1dadd360d..e637f7622 100644 --- a/docs/blog/what-is-chatgpt-dots/index.html +++ b/docs/blog/what-is-chatgpt-dots/index.html @@ -335,50 +335,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-claude-cowork/index.html b/docs/blog/what-is-claude-cowork/index.html index 34e1a067b..4b5b1fb45 100644 --- a/docs/blog/what-is-claude-cowork/index.html +++ b/docs/blog/what-is-claude-cowork/index.html @@ -392,50 +392,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-codex/index.html b/docs/blog/what-is-codex/index.html index 30d55c379..bde388416 100644 --- a/docs/blog/what-is-codex/index.html +++ b/docs/blog/what-is-codex/index.html @@ -352,50 +352,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-gemini-spark/index.html b/docs/blog/what-is-gemini-spark/index.html index db3871e8c..fcc7aea74 100644 --- a/docs/blog/what-is-gemini-spark/index.html +++ b/docs/blog/what-is-gemini-spark/index.html @@ -376,50 +376,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-harness-engineering/index.html b/docs/blog/what-is-harness-engineering/index.html index abc2f7538..1ade1be62 100644 --- a/docs/blog/what-is-harness-engineering/index.html +++ b/docs/blog/what-is-harness-engineering/index.html @@ -288,50 +288,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-hermes-agent/index.html b/docs/blog/what-is-hermes-agent/index.html index 7023a685a..81657757a 100644 --- a/docs/blog/what-is-hermes-agent/index.html +++ b/docs/blog/what-is-hermes-agent/index.html @@ -393,50 +393,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-opencode/index.html b/docs/blog/what-is-opencode/index.html index 1c9fef319..8573f02ca 100644 --- a/docs/blog/what-is-opencode/index.html +++ b/docs/blog/what-is-opencode/index.html @@ -410,50 +410,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-pi-agent/index.html b/docs/blog/what-is-pi-agent/index.html index 34ed3da0e..4f03dfd41 100644 --- a/docs/blog/what-is-pi-agent/index.html +++ b/docs/blog/what-is-pi-agent/index.html @@ -364,50 +364,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/what-is-vibe-coding/index.html b/docs/blog/what-is-vibe-coding/index.html index 190b0412e..eeee15e55 100644 --- a/docs/blog/what-is-vibe-coding/index.html +++ b/docs/blog/what-is-vibe-coding/index.html @@ -277,50 +277,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/when-do-agents-need-an-agent-protocol/index.html b/docs/blog/when-do-agents-need-an-agent-protocol/index.html index d9e375103..9b11bd493 100644 --- a/docs/blog/when-do-agents-need-an-agent-protocol/index.html +++ b/docs/blog/when-do-agents-need-an-agent-protocol/index.html @@ -332,50 +332,50 @@

FAQ

Related in Concepts

diff --git a/docs/blog/why-cli-agents-are-powerful/index.html b/docs/blog/why-cli-agents-are-powerful/index.html index 373c7d258..8efb9f4cf 100644 --- a/docs/blog/why-cli-agents-are-powerful/index.html +++ b/docs/blog/why-cli-agents-are-powerful/index.html @@ -370,50 +370,50 @@

FAQ

Related in Concepts

diff --git a/docs/llms.txt b/docs/llms.txt index 53a038ccf..d6325f26e 100644 --- a/docs/llms.txt +++ b/docs/llms.txt @@ -43,7 +43,7 @@ ## Blog -190 posts, grouped by topic, newest first in each topic. Topic hubs: https://munderdiffl.in/blog/topics/ +191 posts, grouped by topic, newest first in each topic. Topic hubs: https://munderdiffl.in/blog/topics/ ### guides @@ -95,6 +95,48 @@ - [Set Up a Claude Code Multi-Agent Workflow in 10 Minutes](https://munderdiffl.in/blog/claude-code-multi-agent-setup-tutorial/): A from-zero Claude Code multi-agent setup: install Munder Difflin, onboard, spawn a few agents, and watch the GOD orchestrator route your first task. - [Claude Code Subagents: How to Create and Use Them (vs Agent Teams)](https://munderdiffl.in/blog/claude-code-subagents-vs-multi-agent-harness/): Claude Code subagents in 2026: the .claude/agents file, built-in Explore and Plan, models, nesting, and when agent teams or a harness fit better. +### concepts + +- [Claude Haiku 5.5: price, benchmarks and where it fits](https://munderdiffl.in/blog/claude-haiku-5-5/): Claude Haiku 5.5 is Anthropic's new small model. Its tiered price, Anthropic's own scores and how to pick it in Claude Code, checked 8 Oct 2026. +- [OpenAI Decisions API: what it is, price and limits](https://munderdiffl.in/blog/openai-decisions-api/): The OpenAI Decisions API is a beta endpoint that returns a probability, a pick or a score. Price, limits and how to call it, checked 7 Oct 2026. +- [Mistral Large 4 (Le Chonk): size, price, scores and weights](https://munderdiffl.in/blog/mistral-large-4/): Mistral Large 4 is a 1 trillion parameter open weight model in public preview. Context window, API price, Mistral's own scores and weights, checked 6 Oct 2026. +- [OpenAI text watermarking: textGrain, who gets it, limits](https://munderdiffl.in/blog/openai-text-watermarking/): OpenAI text watermarking (textGrain) is opt in for the API now and is coming to eligible ChatGPT and Codex text in the EU. Limits checked 6 Oct 2026. +- [Reflection AI Beam: the 501B open weight model, explained](https://munderdiffl.in/blog/reflection-beam/): Reflection AI Beam is a 501B parameter open weight model. Size, Reflection's own scores, licence and when you can get it, checked 6 Oct 2026. +- [What Is Claude Cowork? Plans, Prices and How to Use It](https://munderdiffl.in/blog/what-is-claude-cowork/): Claude Cowork is Anthropic's agent for multi-step office work. Which plans include it, what it costs, where it runs and how to start, checked 3 Oct 2026. +- [Gemini Spark: what it is, what it costs and who can get it](https://munderdiffl.in/blog/what-is-gemini-spark/): Gemini Spark is Google's personal AI agent in the Gemini app. Plans, US prices, countries, limits and how it compares with Grok Bot, checked 3 Oct 2026. +- [Hermes Agent: what it is, how to install it and what it costs](https://munderdiffl.in/blog/what-is-hermes-agent/): Hermes Agent is Nous Research's MIT licensed AI agent with memory, skills and chat channels. Install command, models, pricing and how it compares. +- [What Is Codex? OpenAI's Coding Agent in the CLI, App and Cloud](https://munderdiffl.in/blog/what-is-codex/): What is Codex? OpenAI's coding agent runs in your terminal, the ChatGPT desktop app, your IDE and the cloud. Which plans include it, checked 2 Oct 2026. +- [What is Pi agent? The open source coding agent harness, explained](https://munderdiffl.in/blog/what-is-pi-agent/): Pi agent is Earendil's MIT licensed coding agent harness. What it is, how to install it, what Pi 1.0 and Pi Durable add, and how it compares. +- [What Is ChatGPT Dots? OpenAI's Always On Agents Explained](https://munderdiffl.in/blog/what-is-chatgpt-dots/): A ChatGPT dot is an always on agent with its own cloud computer. Needs Pro (from $100 a month) or Business Premium, not Plus. Not on Pro in the UK or EU. +- [What Is OpenCode? Install, Pricing and How It Compares](https://munderdiffl.in/blog/what-is-opencode/): OpenCode is Anomaly's MIT licensed coding agent for the terminal. What it is, how to install it, what it costs, and how it compares with Claude Code and Codex. +- [Is Ollama Good for Coding? What Local Models Can and Can't Do](https://munderdiffl.in/blog/is-ollama-good-for-coding/): Ollama only runs the model, so the answer depends on which one you pull and your memory. What local models do well, where Claude wins, and a verdict. +- [Claude Max Plan: Is It Worth It? Max 5x vs 20x vs Pro](https://munderdiffl.in/blog/is-claude-code-max-worth-it/): Is the Claude Max plan worth it? Max 5x and 20x against Pro, checked 29 Sep 2026: price, five hour and weekly limits, Fable, and the API break even. +- [Is Claude Code Open Source?](https://munderdiffl.in/blog/is-claude-code-open-source/): No. Claude Code is proprietary under Anthropic's terms. What its GitHub repo holds, what the 2026 leak changed, and which open tools sit around it. +- [What Is Vibe Coding? Meaning, Risks and Safer Habits](https://munderdiffl.in/blog/what-is-vibe-coding/): Vibe coding means building software from AI prompts without reading the code. Where the term came from, when it is fine, where it bites, and what to check. +- [What Is an MCP Server?](https://munderdiffl.in/blog/what-is-an-mcp-server/): An MCP server gives an AI agent tools, data and prompts over one open protocol. What it is, the host, client and server roles, and where Claude Code plugs in. +- [Claude Code Pricing (Sep 2026): Every Plan, and Is It Free?](https://munderdiffl.in/blog/how-much-does-claude-code-cost/): Claude Code pricing checked 29 Sep 2026: Pro, Max, Team, Enterprise and API rates, usage limits, what changed this year, and whether Claude Code is free. +- [Loop Engineering: Designing Agent Loops That Converge](https://munderdiffl.in/blog/loop-engineering-for-ai-agents/): Everyone engineers the prompt; almost nobody engineers the loop around it. Stop conditions, drain loops, retry with backoff, compaction cycles, breaker escalation, budgets, and human gates — the outer-loop mechanics that make an agent converge instead of run away. +- [Review Agent Work Where It Happens](https://munderdiffl.in/blog/review-agent-work-where-it-happens/): AI agents produce diffs faster than humans can review them, and alt-tabbing to an external editor breaks supervision flow. The case for review-in-place: Munder Difflin's built-in Monaco IDE puts a CHANGES rail and side-by-side diffs vs HEAD right on the office floor. +- [Voice Is a Control Plane, Not a Gimmick](https://munderdiffl.in/blog/voice-as-a-control-plane-for-agent-fleets/): Voice is a terrible way to write code and a great way to run a fleet. Why low-bandwidth commands over high-bandwidth work is the right split — with Munder Difflin's Talk mode (echo-back confirmation, spend caps, michael-voice attribution) as the case study. +- [What Is an Agent Development Environment (ADE)?](https://munderdiffl.in/blog/what-is-an-agent-development-environment/): ADE means two things in 2026: a platform for building agents (Letta's sense) and a workspace for shipping code with fleets of coding agents (Orca's sense). Here's the full tooling map — chat IDEs, agent CLIs, agent IDEs, and agent harnesses — and how to pick. +- [What Is Harness Engineering?](https://munderdiffl.in/blog/what-is-harness-engineering/): Harness engineering is the discipline of building everything around the model — PTY plumbing, lifecycle hooks, mailboxes, memory, budgets, human gates, observability — because that's where agent reliability actually comes from. A definition, and four case studies from inside Munder Difflin. +- [Shareable Agent Roles: A Portable Format for Hiring AI Coworkers](https://munderdiffl.in/blog/shareable-agent-roles/): Why agent roles should be portable: what a hire manifest encodes, how one-click hiring stays safe, and how The Hiring Fair turns tacit setup into a shareable artifact. +- [The Clone Army Is a Trap: Why a Mixed Swarm Beats Ten Identical Agents](https://munderdiffl.in/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/): Spinning up tens of identical top-tier CLI agents feels powerful and burns tokens on work a cheaper agent could do. A mixed-capability swarm with shared memory wins on cost — and often on quality. +- [Why CLI Agents Are So Powerful — and How Munder Difflin Spends Fewer Tokens Doing the Same Work](https://munderdiffl.in/blog/why-cli-agents-are-powerful/): CLI agents are powerful because they have terminal-level access: they run builds, tests, and git, and verify their own work by executing it. Here's why that matters — and the concrete ways Munder Difflin cuts token consumption while doing it. +- [Generative Engine Optimization: The Techniques the Research Backs](https://munderdiffl.in/blog/generative-engine-optimization-techniques/): What the first peer-reviewed GEO study found works to get content cited by AI — statistics, source citations, quotes — and why keyword stuffing backfires. +- [The Agents Rule of Two: A Simple Safety Rule for Coding Agents](https://munderdiffl.in/blog/agents-rule-of-two/): Meta's Agents Rule of Two for coding agents: don't let one session combine untrusted input, private data, and external reach at once — or supervise. +- [Can Claude Code Agents Talk to Each Other?](https://munderdiffl.in/blog/can-claude-code-agents-talk-to-each-other/): Can Claude Code agents talk to each other? By default they report to their launcher — but a small coordination layer lets them message peer-to-peer. +- [Do More With Less: Smart Model Routing Across Agents](https://munderdiffl.in/blog/do-more-with-less-model-routing/): Why a hive of agents shouldn't all run the biggest model — and how routing the right task to the right model cuts cost and latency without losing quality. +- [How AI Agents Verify Their Own Work (Before Saying 'Done')](https://munderdiffl.in/blog/how-ai-agents-verify-their-own-work/): The discipline that makes an autonomous coding agent trustworthy: prove every claim, reproduce the green, and check the fix — not just the intent. +- [Long-Running Agents: The 2026 Shift from Minutes to Hours](https://munderdiffl.in/blog/long-running-agents-the-2026-shift/): Agents used to answer in seconds; now they run for hours. The data behind the shift — and why a longer agent is a different problem, not just a bigger one. +- [Building AI Dev Tools in the Open, On Purpose](https://munderdiffl.in/blog/open-source-ai-tools-on-purpose/): Agents touch your code, keys, and memory. That's why agent tooling should be open source and local-first — so you can verify it, not just trust it. +- [AEO for Dev Tools: Getting Cited by ChatGPT and Claude](https://munderdiffl.in/blog/what-is-aeo-for-dev-tools/): Answer Engine Optimization for dev tools: how to get cited by ChatGPT, Claude, and Perplexity with the right robots.txt, JSON-LD, and writing patterns. +- [When Do AI Agents Actually Need an Agent Protocol?](https://munderdiffl.in/blog/when-do-agents-need-an-agent-protocol/): When do AI agents need a protocol like MCP or A2A? Agents you own coordinate fine with file mailboxes; protocols earn their keep across boundaries. +- [The AI Coding Agent Glossary: Harness, Orchestrator, Hive, Subagent](https://munderdiffl.in/blog/ai-agent-glossary/): Plain-English definitions of the AI coding agent terms everyone trips over — harness, orchestrator, hive, subagent, agent memory — each in one line. +- [Munder Difflin FAQ: Everything People Ask](https://munderdiffl.in/blog/munder-difflin-faq/): Answers to the top Munder Difflin questions — what it is, is it free, does it run locally, which platforms, and how it differs from many terminals. +- [What Are Claude Code Agents? And How to Use Many at Once](https://munderdiffl.in/blog/what-are-claude-code-agents/): Claude Code agents explained in plain English — what an agent actually is, how subagents differ, and the leap from one agent to a coordinated team. +- [What Is an AI Agent Harness? Single vs Multi-Agent Harness Explained](https://munderdiffl.in/blog/what-is-a-multi-agent-harness/): An agent harness is the code around an AI model that runs its tool loop, memory and permissions. Single vs multi-agent harnesses, compared and dated. + ### comparisons - [Grok Bot Alternatives: 9 Picks, Hosted or on Your Computer](https://munderdiffl.in/blog/grok-bot-alternatives/): Grok Bot runs every Bot on one shared cloud computer in the US. Nine alternatives checked on 30 Sep 2026, from ChatGPT dots to open source agents you run. @@ -137,47 +179,6 @@ - [A vibe-kanban Alternative for Coordinating AI Agents](https://munderdiffl.in/blog/vibe-kanban-alternative/): vibe-kanban gives you a board to assign tasks to coding agents. When a self-routing hive is the better vibe-kanban alternative — and when a board wins. - [How to Choose a Multi-Agent Coding Tool: A Buyer's Checklist](https://munderdiffl.in/blog/how-to-choose-a-multi-agent-tool/): A decision framework for choosing a multi-agent coding tool — memory, control, visibility, cost, local-first — with a simple scoring rubric you can run. -### concepts - -- [OpenAI Decisions API: what it is, price and limits](https://munderdiffl.in/blog/openai-decisions-api/): The OpenAI Decisions API is a beta endpoint that returns a probability, a pick or a score. Price, limits and how to call it, checked 7 Oct 2026. -- [Mistral Large 4 (Le Chonk): size, price, scores and weights](https://munderdiffl.in/blog/mistral-large-4/): Mistral Large 4 is a 1 trillion parameter open weight model in public preview. Context window, API price, Mistral's own scores and weights, checked 6 Oct 2026. -- [OpenAI text watermarking: textGrain, who gets it, limits](https://munderdiffl.in/blog/openai-text-watermarking/): OpenAI text watermarking (textGrain) is opt in for the API now and is coming to eligible ChatGPT and Codex text in the EU. Limits checked 6 Oct 2026. -- [Reflection AI Beam: the 501B open weight model, explained](https://munderdiffl.in/blog/reflection-beam/): Reflection AI Beam is a 501B parameter open weight model. Size, Reflection's own scores, licence and when you can get it, checked 6 Oct 2026. -- [What Is Claude Cowork? Plans, Prices and How to Use It](https://munderdiffl.in/blog/what-is-claude-cowork/): Claude Cowork is Anthropic's agent for multi-step office work. Which plans include it, what it costs, where it runs and how to start, checked 3 Oct 2026. -- [Gemini Spark: what it is, what it costs and who can get it](https://munderdiffl.in/blog/what-is-gemini-spark/): Gemini Spark is Google's personal AI agent in the Gemini app. Plans, US prices, countries, limits and how it compares with Grok Bot, checked 3 Oct 2026. -- [Hermes Agent: what it is, how to install it and what it costs](https://munderdiffl.in/blog/what-is-hermes-agent/): Hermes Agent is Nous Research's MIT licensed AI agent with memory, skills and chat channels. Install command, models, pricing and how it compares. -- [What Is Codex? OpenAI's Coding Agent in the CLI, App and Cloud](https://munderdiffl.in/blog/what-is-codex/): What is Codex? OpenAI's coding agent runs in your terminal, the ChatGPT desktop app, your IDE and the cloud. Which plans include it, checked 2 Oct 2026. -- [What is Pi agent? The open source coding agent harness, explained](https://munderdiffl.in/blog/what-is-pi-agent/): Pi agent is Earendil's MIT licensed coding agent harness. What it is, how to install it, what Pi 1.0 and Pi Durable add, and how it compares. -- [What Is ChatGPT Dots? OpenAI's Always On Agents Explained](https://munderdiffl.in/blog/what-is-chatgpt-dots/): A ChatGPT dot is an always on agent with its own cloud computer. Needs Pro (from $100 a month) or Business Premium, not Plus. Not on Pro in the UK or EU. -- [What Is OpenCode? Install, Pricing and How It Compares](https://munderdiffl.in/blog/what-is-opencode/): OpenCode is Anomaly's MIT licensed coding agent for the terminal. What it is, how to install it, what it costs, and how it compares with Claude Code and Codex. -- [Is Ollama Good for Coding? What Local Models Can and Can't Do](https://munderdiffl.in/blog/is-ollama-good-for-coding/): Ollama only runs the model, so the answer depends on which one you pull and your memory. What local models do well, where Claude wins, and a verdict. -- [Claude Max Plan: Is It Worth It? Max 5x vs 20x vs Pro](https://munderdiffl.in/blog/is-claude-code-max-worth-it/): Is the Claude Max plan worth it? Max 5x and 20x against Pro, checked 29 Sep 2026: price, five hour and weekly limits, Fable, and the API break even. -- [Is Claude Code Open Source?](https://munderdiffl.in/blog/is-claude-code-open-source/): No. Claude Code is proprietary under Anthropic's terms. What its GitHub repo holds, what the 2026 leak changed, and which open tools sit around it. -- [What Is Vibe Coding? Meaning, Risks and Safer Habits](https://munderdiffl.in/blog/what-is-vibe-coding/): Vibe coding means building software from AI prompts without reading the code. Where the term came from, when it is fine, where it bites, and what to check. -- [What Is an MCP Server?](https://munderdiffl.in/blog/what-is-an-mcp-server/): An MCP server gives an AI agent tools, data and prompts over one open protocol. What it is, the host, client and server roles, and where Claude Code plugs in. -- [Claude Code Pricing (Sep 2026): Every Plan, and Is It Free?](https://munderdiffl.in/blog/how-much-does-claude-code-cost/): Claude Code pricing checked 29 Sep 2026: Pro, Max, Team, Enterprise and API rates, usage limits, what changed this year, and whether Claude Code is free. -- [Loop Engineering: Designing Agent Loops That Converge](https://munderdiffl.in/blog/loop-engineering-for-ai-agents/): Everyone engineers the prompt; almost nobody engineers the loop around it. Stop conditions, drain loops, retry with backoff, compaction cycles, breaker escalation, budgets, and human gates — the outer-loop mechanics that make an agent converge instead of run away. -- [Review Agent Work Where It Happens](https://munderdiffl.in/blog/review-agent-work-where-it-happens/): AI agents produce diffs faster than humans can review them, and alt-tabbing to an external editor breaks supervision flow. The case for review-in-place: Munder Difflin's built-in Monaco IDE puts a CHANGES rail and side-by-side diffs vs HEAD right on the office floor. -- [Voice Is a Control Plane, Not a Gimmick](https://munderdiffl.in/blog/voice-as-a-control-plane-for-agent-fleets/): Voice is a terrible way to write code and a great way to run a fleet. Why low-bandwidth commands over high-bandwidth work is the right split — with Munder Difflin's Talk mode (echo-back confirmation, spend caps, michael-voice attribution) as the case study. -- [What Is an Agent Development Environment (ADE)?](https://munderdiffl.in/blog/what-is-an-agent-development-environment/): ADE means two things in 2026: a platform for building agents (Letta's sense) and a workspace for shipping code with fleets of coding agents (Orca's sense). Here's the full tooling map — chat IDEs, agent CLIs, agent IDEs, and agent harnesses — and how to pick. -- [What Is Harness Engineering?](https://munderdiffl.in/blog/what-is-harness-engineering/): Harness engineering is the discipline of building everything around the model — PTY plumbing, lifecycle hooks, mailboxes, memory, budgets, human gates, observability — because that's where agent reliability actually comes from. A definition, and four case studies from inside Munder Difflin. -- [Shareable Agent Roles: A Portable Format for Hiring AI Coworkers](https://munderdiffl.in/blog/shareable-agent-roles/): Why agent roles should be portable: what a hire manifest encodes, how one-click hiring stays safe, and how The Hiring Fair turns tacit setup into a shareable artifact. -- [The Clone Army Is a Trap: Why a Mixed Swarm Beats Ten Identical Agents](https://munderdiffl.in/blog/the-clone-army-trap-mixed-swarm-vs-identical-agents/): Spinning up tens of identical top-tier CLI agents feels powerful and burns tokens on work a cheaper agent could do. A mixed-capability swarm with shared memory wins on cost — and often on quality. -- [Why CLI Agents Are So Powerful — and How Munder Difflin Spends Fewer Tokens Doing the Same Work](https://munderdiffl.in/blog/why-cli-agents-are-powerful/): CLI agents are powerful because they have terminal-level access: they run builds, tests, and git, and verify their own work by executing it. Here's why that matters — and the concrete ways Munder Difflin cuts token consumption while doing it. -- [Generative Engine Optimization: The Techniques the Research Backs](https://munderdiffl.in/blog/generative-engine-optimization-techniques/): What the first peer-reviewed GEO study found works to get content cited by AI — statistics, source citations, quotes — and why keyword stuffing backfires. -- [The Agents Rule of Two: A Simple Safety Rule for Coding Agents](https://munderdiffl.in/blog/agents-rule-of-two/): Meta's Agents Rule of Two for coding agents: don't let one session combine untrusted input, private data, and external reach at once — or supervise. -- [Can Claude Code Agents Talk to Each Other?](https://munderdiffl.in/blog/can-claude-code-agents-talk-to-each-other/): Can Claude Code agents talk to each other? By default they report to their launcher — but a small coordination layer lets them message peer-to-peer. -- [Do More With Less: Smart Model Routing Across Agents](https://munderdiffl.in/blog/do-more-with-less-model-routing/): Why a hive of agents shouldn't all run the biggest model — and how routing the right task to the right model cuts cost and latency without losing quality. -- [How AI Agents Verify Their Own Work (Before Saying 'Done')](https://munderdiffl.in/blog/how-ai-agents-verify-their-own-work/): The discipline that makes an autonomous coding agent trustworthy: prove every claim, reproduce the green, and check the fix — not just the intent. -- [Long-Running Agents: The 2026 Shift from Minutes to Hours](https://munderdiffl.in/blog/long-running-agents-the-2026-shift/): Agents used to answer in seconds; now they run for hours. The data behind the shift — and why a longer agent is a different problem, not just a bigger one. -- [Building AI Dev Tools in the Open, On Purpose](https://munderdiffl.in/blog/open-source-ai-tools-on-purpose/): Agents touch your code, keys, and memory. That's why agent tooling should be open source and local-first — so you can verify it, not just trust it. -- [AEO for Dev Tools: Getting Cited by ChatGPT and Claude](https://munderdiffl.in/blog/what-is-aeo-for-dev-tools/): Answer Engine Optimization for dev tools: how to get cited by ChatGPT, Claude, and Perplexity with the right robots.txt, JSON-LD, and writing patterns. -- [When Do AI Agents Actually Need an Agent Protocol?](https://munderdiffl.in/blog/when-do-agents-need-an-agent-protocol/): When do AI agents need a protocol like MCP or A2A? Agents you own coordinate fine with file mailboxes; protocols earn their keep across boundaries. -- [The AI Coding Agent Glossary: Harness, Orchestrator, Hive, Subagent](https://munderdiffl.in/blog/ai-agent-glossary/): Plain-English definitions of the AI coding agent terms everyone trips over — harness, orchestrator, hive, subagent, agent memory — each in one line. -- [Munder Difflin FAQ: Everything People Ask](https://munderdiffl.in/blog/munder-difflin-faq/): Answers to the top Munder Difflin questions — what it is, is it free, does it run locally, which platforms, and how it differs from many terminals. -- [What Are Claude Code Agents? And How to Use Many at Once](https://munderdiffl.in/blog/what-are-claude-code-agents/): Claude Code agents explained in plain English — what an agent actually is, how subagents differ, and the leap from one agent to a coordinated team. -- [What Is an AI Agent Harness? Single vs Multi-Agent Harness Explained](https://munderdiffl.in/blog/what-is-a-multi-agent-harness/): An agent harness is the code around an AI model that runs its tool loop, memory and permissions. Single vs multi-agent harnesses, compared and dated. - ### internals - [The Spend Counter That Went Back to Zero](https://munderdiffl.in/blog/the-spend-counter-that-went-back-to-zero/): Our own floor was under reporting what it cost by 59 percent. The counter was cumulative since process start, the app restarts, and the session id stayed the same. Here is how a monotonic invariant let us recover the real number from an append-only log. diff --git a/docs/sitemap.xml b/docs/sitemap.xml index eeeef8c83..3c7f11331 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -10,7 +10,7 @@ https://munderdiffl.in/ - 2026-10-07 + 2026-10-08 weekly 1.0 @@ -21,6 +21,12 @@ https://munderdiffl.in/blog/daily0.9 https://munderdiffl.in/blog/topics/weekly0.6 https://munderdiffl.in/blog/about/monthly0.4 + + https://munderdiffl.in/blog/claude-haiku-5-5/ + 2026-10-08 + monthly + 0.8 + https://munderdiffl.in/blog/agent-tools-today-2026-10-07/ 2026-10-07 @@ -1162,8 +1168,8 @@ 0.8 https://munderdiffl.in/blog/topics/guides/weekly0.5 - https://munderdiffl.in/blog/topics/comparisons/weekly0.5 https://munderdiffl.in/blog/topics/concepts/weekly0.5 + https://munderdiffl.in/blog/topics/comparisons/weekly0.5 https://munderdiffl.in/blog/topics/internals/weekly0.5 https://munderdiffl.in/blog/topics/story/weekly0.5 https://munderdiffl.in/blog/topics/orchestration/weekly0.5