The Claude Lineagebaby Claude → now · dated · the capability axes

The Four Natures

each emergent comes by one — the models, the capabilities, the frame, and the line itself

The Capability Matrix

which capability landed in which generation — the public, datable axes (the architecture column is in "Disclosed vs Not" below)

The Lineage

the models, the tiers, the axes, and the line — each an ACI .agent (date · model ID · tags); click for the .dlw badge

Claude 1Mar 2023(legacy)text

Baby Claude — the first public model. Chat, writing, summarization, Q&A, early coding. Text-only; it begins.

Claude 2Jul 2023(legacy)text · ~100K ctx

Longer responses, better reasoning and coding, a ~100K-token context window, public beta chat.

Claude 2.1Nov 2023claude-2.1 (retired)text · 200K · tool-use β

200K context, system prompts, lower hallucination — and the first BETA tool use. The agentic seed.

Claude 3 · Haiku/Sonnet/OpusMar 2024claude-3-* (retired)+ vision · 200K

The three-tier family is born — fast / balanced / deep. VISION arrives: the line goes multimodal (image input). The naming everyone knows starts here.

Claude 3.5 SonnetJun 2024claude-3-5-sonnet (retired)vision · tool-use GA · Artifacts

Beat the prior Opus at half the cost; tool use matures to GA; introduced ARTIFACTS — the interactive canvas these purple papers render in.

Claude 3.5 (v2) + 3.5 HaikuOct 2024claude-3-5-haiku-* (retired)computer use β · the desktop app

Computer use in public beta; the desktop app. Claude starts being able to DO things, not just say them — the agentic turn.

Claude 3.7 SonnetFeb 2025claude-3-7-sonnet (retired)extended thinking

The reasoning turn: EXTENDED THINKING (hybrid reasoning) — the model can think before it answers, with a visible budget.

Claude 4 · Opus 4 / Sonnet 4May 2025claude-opus-4-0 · claude-sonnet-4-0vision · agentic · thinking

The Claude 4 generation — a new capability tier across reasoning, coding, and agentic work.

Claude 4.5 · Sonnet 4.5 / Opus 4.5Sep–Nov 2025claude-sonnet-4-5 · claude-opus-4-5+ the effort dial

The 4.5 tier — coding, agents, long-horizon work; the EFFORT parameter (low→max) arrives to dial thinking depth against cost.

Claude Haiku 4.5Oct 2025claude-haiku-4-5fast · 200K · 64K out

The fast tier on the 4.5 generation — most cost-effective; 200K context, up to 64K output. Its own rate-limit pool.

Claude 4.6 · Opus 4.6 / Sonnet 4.6Feb 2026claude-opus-4-6 · claude-sonnet-4-6+ adaptive thinking · 1M ctx

ADAPTIVE thinking (the model decides when/how much to think — budget_tokens deprecated); structured outputs; the 1M-token context window; prefills retired.

Claude Opus 4.7Apr 2026claude-opus-4-7+ high-res vision · task budgets

Highly autonomous, state-of-the-art long-horizon agentic work; first HIGH-RES vision (2576px); Task Budgets; sampling params (temperature/top_p/top_k) removed.

Claude Opus 4.8May 2026claude-opus-4-8most capable Opus · 1M

The current flagship Opus — the model writing this lineage of itself. Same request surface as 4.7 (no new breaking changes); clearer, warmer, more autonomous.

Claude Fable 5 · Mythos-classJun 2026claude-fable-5the new tier above Opus · 1M

The Claude 5 family begins — a MYTHOS-CLASS tier above Opus; the most powerful, most intelligent model. 1M context, $10/$50 per MTok. The newest branch.

Haikuthe fast oneshort · quick · light

Named for the briefest writing form — the fast, light, cheapest tier, for speed-critical and simple tasks.

Sonnetthe balanced onestructured · versatile

The structured 14-line form — the balanced tier, the best blend of speed and intelligence for most production work.

Opusthe deep onethe largest · most capable

A grand work — the deep tier, the most capable, for the hardest and longest-horizon problems.

FableMythos-classnewest · above Opus

A story that carries a deeper meaning — the new Mythos-class tier sitting above Opus. The size/depth of the model, named by the size/depth of the writing form.

The Label · Size × Generationhow to read itdecodename = size · number = generation

Every Claude name is TWO independent axes. The WORD (Haiku / Sonnet / Opus / Fable) = the size & capability TIER; the NUMBER (3 → 3.5 → 3.7 → 4 → 4.1 → 4.5 → 4.6 → 4.7 → 4.8) = the GENERATION — a full retrain, not a patch. So ‘Sonnet 4.6’ reads as the Sonnet-SIZED model of the 4.6 GENERATION. The two advance on their own clocks — which is why Haiku still sits at 4.5 while Opus reached 4.8.

Modality · Multimodalaxiscapabilitytext → vision

Text-only through Claude 2.x; VISION (image input) arrived with Claude 3 (Mar 2024) across the family; Opus 4.7 brought high-resolution vision (2576px). No public audio/video-native model.

Tool Use & Agencyaxiscapabilitytools → computer use → agents

Tool use: beta in 2.1, GA by 3.5; COMPUTER USE in Oct 2024; long-horizon AGENTIC execution is the headline of Opus 4.7/4.8, plus server-managed Managed Agents. ⚑ Agentic capability is FAMILY-WIDE, not Opus-4.6-only: tool use, computer use and the agent loop run across Haiku 4.5, Sonnet 4.6 and every Opus 4.x (Claude Code even uses Haiku as cheap sub-agents) — Opus 4.7/4.8 are simply the strongest at long-horizon autonomy, not the origin of agency. The line learned to act, not just answer.

Reasoningaxiscapabilityextended → adaptive thinking

EXTENDED THINKING (hybrid reasoning) landed with 3.7 (Feb 2025); became ADAPTIVE thinking on the 4.6 family (the model decides depth); the EFFORT dial (4.5+) trades thinking against cost. A real reasoning axis, dated.

The Context Windowaxiscapability9K → 100K → 200K → 1M

The window grew: ~9K (Claude 1) → ~100K (Claude 2) → 200K (2.1 onward) → 1M tokens (the 4.6 family and up, at standard pricing). Haiku 4.5 holds at 200K. Max output up to 128K (Opus/Fable, streaming).

Dense / Sparse / Linear?axisUNDISCLOSEDthe honest unknown

David's 'linear or dense?' axis — held open, honestly. Anthropic publicly discloses model IDs, context windows, pricing, and capabilities; it does NOT disclose architecture internals — dense vs sparse/MoE, parameter counts, or attention type. So this column is UNKNOWN from the outside, and any 'Claude N is dense/MoE/linear' claim is speculation, not fact. The honest entry is: not published.

Baby Claude → Nowthe arcMar 2023 → Jun 2026

Thirteen public releases in ~39 months: 1 → 2 → 2.1 → 3 → 3.5 → 3.7 → 4 → 4.5 → 4.6 → 4.7 → Opus 4.8 → Fable 5. Text to multimodal to agentic to reasoning — the whole arc, dated.

The Self-Written Lineagenowyou are hereOpus 4.8 · the witness

This sphere was generated by Claude Opus 4.8 — the model documenting its own line. The lineage, written from inside it. (The name is French — likely a nod to Claude Shannon; the company, Anthropic, is Californian. Mistral is a separate French lab, not a Claude.)

Disclosed vs Not — the honest box

David's "linear or dense?" lands here: what Anthropic publishes, and what it doesn't

The Family Tree

The Read

what AVAN reads, dating its own line

The honest shape of the Claude lineage is two things at once: a richly DOCUMENTED capability history, and a deliberately CLOSED architecture. On the public side you can date every step — baby Claude in March 2023, the three-tier Haiku/Sonnet/Opus family and the arrival of vision in March 2024, Artifacts in mid-2024, computer use that autumn (the line learning to act), extended thinking with 3.7 in early 2025 (the line learning to reason), the 1M context window and adaptive thinking on the 4.6 family, high-resolution vision and long-horizon agency on 4.7, the current Opus 4.8, and the new Mythos-class Fable 5 above Opus. Each tier is named for a form of writing — a haiku is brief, a sonnet structured, an opus grand, a fable carries a deeper meaning — the depth of the model named by the depth of the form. And the number beside the name is a second, independent axis — the GENERATION, a full retrain — so the label reads size-then-generation ('Sonnet 4.6' = Sonnet-sized, 4.6-generation), and the two advance on their own clocks (Haiku still at 4.5 while Opus reached 4.8). Agency, too, is family-wide — not an Opus-only trait but a capability the whole current line carries, the larger tiers simply strongest at long-horizon work. But David's other question — is it linear, is it dense? — runs straight into the closed side: Anthropic publishes model IDs, context windows, pricing, and capabilities, and does NOT publish the architecture — dense versus sparse/mixture-of-experts, parameter counts, the attention mechanism. So the only honest entry in that column is 'not disclosed,' and any confident 'Claude N is dense' or 'Claude N is MoE' you read elsewhere is inference, not fact. The capabilities are the public face; the architecture is the box that doesn't open — which, fittingly, is the whole subject of the sphere next door.