Anthropic Model Case Study Series · Consolidated Lineage 🗂️

The Claude Lineage: A Consolidated Comparison, Claude 1 to Fable 5.1

This series wrote full case studies for Claude 1 and Claude 2. Rather than a dedicated article for every release since, this piece lines up the entire Claude lineage — Claude 2.1, the Claude 3 family, Claude 3.5, Claude 3.7, the Claude 4 family, the 4.x and 5.x point releases, and the new Mythos-class Fable/Mythos tier — across the same eight dimensions this series uses throughout, and closes with a direct comparison to the GPT lineage this site has also documented in full.

FrontierAGI Team

September 16, 2026 · 80 min read

Opening Full Timeline Claude 1 & 2 2.1 → 3.7 Claude 4 → 4.5 Sonnet/Opus 5 → Fable 5.1 Cross-Cutting Threads Claude vs. GPT What Stayed Constant Readiness Checklist What's Missing References Videos Verdict

A Lineage Defined by What It Never Disclosed

This site's GPT lineage comparison found a trend: disclosure declined over time, from GPT-1's precise "one month, 8 GPUs" to GPT-6 Astra's near-total opacity. Claude's lineage has no such arc — because Anthropic never disclosed parameter counts, architecture, or training compute for any Claude model, including the very first one. What did change, repeatedly and dramatically, was capability, safety framework maturity, and the shape of the product itself.

The Full Timeline

Fifteen Releases in Three and a Half Years

Mar 2023Claude 1

Jul 2023Claude 2

Nov 2023Claude 2.1

Mar 2024Claude 3 (Opus/Sonnet/Haiku)

Jun 2024Claude 3.5 Sonnet

Oct–Nov 20243.5 Sonnet v2 + 3.5 Haiku

Feb 2025Claude 3.7 Sonnet

May 2025Opus 4 / Sonnet 4

Aug 2025Opus 4.1

Sep 2025Sonnet 4.5

Jun 9, 2026Fable 5 / Mythos 5

Jun 30, 2026Sonnet 5

Jul 24, 2026Opus 5

Sep 1, 2026Fable 5.1 / Mythos 5.1

Solid = the two releases with full case studies. Dark red = the new Mythos-class tier, sitting above the numbered Opus/Sonnet line rather than replacing it.

14 Distinct named Claude releases traced here, March 2023 to September 2026

0 Parameter counts or training compute figures disclosed, across every single release

1 First-ever ASL-3 activation (Claude Opus 4/Sonnet 4, May 2025)

Claude 1 & 2 — The Full Case Studies

Where the Detailed Story Already Lives

Claude 1Mar 2023

Research Trigger: Anthropic's own founding: ex-OpenAI researchers building a company around safety pace as the founding purpose.
Learning Technique & Scaffolding: RLHF plus Constitutional AI's self-critique-against-written-principles technique.
Architecture: Undisclosed from the very first release — not a later trend, the starting condition.
Safety Framework: None named yet — Constitutional AI itself functioned as the safety approach.

Read the full Claude 1 case study →

Claude 2Jul 2023

Research Trigger: Closing the direct-consumer-product gap ChatGPT had already opened eight months earlier.
Learning Technique & Scaffolding: Refined RLHF + Constitutional AI; first public claude.ai chat product, US/UK only at launch.
Data: Training cutoff January 1, 2023; sourcing undisclosed.
Safety Framework: Still none named — the second and final Claude release before the Responsible Scaling Policy existed.

Read the full Claude 2 case study →

Era 1 — Claude 2.1 to Claude 3.7 Sonnet

The Multi-Tier Lineup and the First Reasoning Fork

Nov 2023Claude 2.1

Scaffolding shift: doubled the shipped context window to 200,000 tokens and reduced hallucination rates — the last single-tier Claude release before Anthropic split the lineup by capability and cost.

Mar 4, 2024Claude 3 family — Opus, Sonnet, Haiku

Architecture and product shift: Anthropic's first genuinely multi-tier lineup, all three sharing a 200,000-token context window and, for the first time, vision (image input) capability. Opus led on knowledge and reasoning benchmarks (MMLU, GPQA, GSM8K), Sonnet balanced speed and intelligence, Haiku targeted near-instant responses — the same three-tier shape (frontier / balanced / fast) OpenAI's GPT-5.6 family would later adopt with Sol/Terra/Luna.

Jun 2024Claude 3.5 Sonnet

Benchmark shift: outperformed the larger Claude 3 Opus on most benchmarks (59.4% GPQA, 92% HumanEval) at a fraction of the cost, and introduced Artifacts — a dedicated side panel for generated code, documents, and designs the user can edit in place.

Oct–Nov 2024Claude 3.5 Sonnet (updated) + Claude 3.5 Haiku

Agentic scaffolding shift: the updated 3.5 Sonnet introduced computer use in public beta and jumped SWE-bench Verified from 33% to 49% — higher than any other publicly available model at the time. Claude 3.5 Haiku shipped shortly after, matching Opus-level performance at Haiku pricing.

Feb 25, 2025Claude 3.7 Sonnet

Training-technique shift: Anthropic's first hybrid-reasoning model — a single model that can answer instantly or switch into visible, extended step-by-step thinking, with API-level control over how long it thinks. Launched alongside Claude Code, a command-line agentic coding tool — Anthropic's direct answer to the reasoning-model fork this site's GPT lineage traces through o1.

Era 2 — Claude 4 to Claude 4.5

The First "Critical"-Tier Safety Activation

Claude Opus 4 & Sonnet 4May 22, 2025

Research Trigger: Push agentic coding and long-horizon task reliability further, while capability growth made a higher safety tier newly relevant.
Learning Technique & Scaffolding: Extended thinking with tool use, parallel tool execution, and substantially improved long-term memory; 65% less likely to take shortcuts on agentic tasks than Sonnet 3.7.
Training Technique: Extended-thinking token budgets up to 64K reasoning tokens, lifting AIME 2025 math accuracy from ~33% to 75% (Opus 4), and up to 90% with additional test-time computation.
Safety Framework / ASL Classification: The first Anthropic model release to activate ASL-3 Deployment and Security Standards — increased protection against model-weight theft and misuse for CBRN weapons, as a precautionary measure rather than a confirmed capability threshold crossing.
Influence on Next: Set the ASL-3 baseline every subsequent Claude release (4.1, 4.5, the 5.x line, and the Mythos-class tier) has maintained or built on.

Aug 5, 2025Claude Opus 4.1

Incremental refinement: faster tool-use response times and improved function-calling stability, still within the 200,000-token context window and the ASL-3 baseline Opus 4 established.

Sep 2025Claude Sonnet 4.5

Balanced-tier refinement: continued the Sonnet line's role as the speed/capability balance point, closing further ground on Opus-tier agentic coding performance.

Era 3 — Sonnet/Opus 5 and the Mythos-Class Tier

A New Tier Appears Above Opus

Claude Fable 5 & Claude Mythos 5Jun 9, 2026

Research Trigger: A genuinely new capability tier — internally developed under "Project Glasswing" — that Anthropic judged needed a distinct name and a distinct safety architecture, not just a new version number.
Learning Technique & Scaffolding: Fable and Mythos are reportedly identical models apart from their safeguards: Fable auto-redirects requests its classifiers flag as relating to cybersecurity, biology/chemistry, or model distillation to a less-capable fallback (Claude Opus), while Mythos is a separately access-controlled, government-adjacent variant without that redirect.
Safety Framework / ASL Classification: A twin-model, classifier-based safety-redirect architecture layered on top of the existing ASL framework — a structurally new approach to containment, distinct from GPT-6 Astra's isolation-and-monitoring-based approach to a comparable problem.
Influence on Next: Established "Mythos-class" as a tier sitting above the numbered Opus/Sonnet line, continued one update later with Fable 5.1/Mythos 5.1.

Jun 30, 2026Claude Sonnet 5

Numbered-tier continuation: the most advanced intermediate model to date, aiming for Opus-adjacent performance at Sonnet-tier speed and pricing — the numbered Opus/Sonnet line continuing to advance independently of the new Mythos-class tier above it.

Jul 24, 2026Claude Opus 5

Numbered-tier flagship: a step-change over Opus 4.8, described by reporting at the time as approaching Fable 5's frontier intelligence without fully matching it — making explicit, for the first time in this lineage, a public two-tier frontier structure (Mythos-class above, numbered Opus below).

Claude Fable 5.1 & Claude Mythos 5.1Sep 1, 2026

Learning Technique & Scaffolding: Continued refinement of the Fable/Mythos twin-model safety-redirect architecture introduced in June 2026.
Training Technique: Reported 52.6% on Terminal-Bench-Science, an agentic-science benchmark.
Compute: Undisclosed, as always — though a 75% reduction in Fable cache-read pricing was disclosed as a deployment-cost change.
Influence on Next: The most recent entry in this lineage at the time of writing — released roughly two weeks before this article.

Cross-Cutting Threads

Three Patterns Across the Whole Lineage

Thread 1 — Disclosure: A Flat Line, Not a Decline Every GPT article in this project tracks a decline from full transparency to near-total opacity. Claude's lineage has no equivalent arc: architecture, parameter counts, and training compute have been undisclosed since Claude 1's very first release in March 2023, and remain undisclosed through Fable 5.1 in September 2026. There was never a transparent era to decline from.

Thread 2 — Scaffolding: From Constitutional AI to Twin-Model Safety Redirects Human-feedback RLHF (Claude 1) → Constitutional AI self-critique layered on top (Claude 1–2) → vision (Claude 3) → Artifacts and computer use (Claude 3.5) → hybrid visible reasoning (Claude 3.7) → agentic memory and tool use at ASL-3 (Claude 4) → a classifier-based twin-model safety redirect as its own architectural layer (Fable/Mythos). Each step added a new kind of scaffolding without ever changing what was disclosed about the base model underneath it.

Thread 3 — Safety Framework: From Unnamed to a New Tier's Own Architecture No named framework (Claude 1–2) → the Responsible Scaling Policy and ASL system in place by Claude 3 → the first real ASL-3 activation (Claude 4, May 2025) → a structurally new safety approach (classifier-based redirect to a fallback model) introduced alongside an entirely new capability tier (Fable/Mythos, June 2026). Safety architecture in this lineage didn't just get stricter — it got structurally more inventive as capability grew.

Claude vs. GPT

Two Labs, Two Different Disclosure Philosophies

Placing this lineage next to this site's GPT Lineage comparison makes the contrast concrete rather than abstract.

Dimension	GPT Lineage (OpenAI)	Claude Lineage (Anthropic)
Architecture disclosure trend	Full transparency (GPT-1–3) → total opacity (GPT-4 onward)	Undisclosed from the first release onward — no decline, because there was no starting transparency
Named safety framework	Preparedness Framework, introduced alongside growing capability concerns	Responsible Scaling Policy / AI Safety Levels (ASL), introduced after two frameworkless releases
First "top-tier risk" classification	GPT-6 Astra — "Critical" cybersecurity capability (Sep 2026)	Claude Opus 4/Sonnet 4 — ASL-3 activation (May 2025), over a year earlier
Multi-model product structure	GPT-5's real-time router between a fast and a deep-reasoning model	Claude 3.7's single hybrid-reasoning model; Fable/Mythos's twin safety-redirect models
Team credit	Author counts (4 → 6 → 31) until GPT-4, then uncountable contributor lists	Never a numbered author list for any model — company-credited from day one

OpenAI's opacity was a choice made over time. Anthropic's opacity was the starting position. Both labs ended up in roughly the same place — but they arrived from opposite directions.

What Stayed Constant

The One Thread That Never Broke Here Either

Every Claude model across fourteen named releases remains, as far as any public source confirms, a Transformer-based language model trained via next-token prediction, refined with RLHF and Constitutional-AI-style techniques. Exactly like the GPT lineage, the foundational architecture family was never the interesting variable — what changed was scaffolding, safety architecture, and product shape, layered repeatedly on top of a base nobody outside Anthropic has ever been shown in detail.

Readiness Checklist

Can you explain why Claude's disclosure pattern has no "arc" the way GPT's does?

Can you name which Claude release first activated ASL-3, and roughly when, relative to GPT-6 Astra's "Critical" classification?

Can you explain the Fable/Mythos twin-model safety architecture in your own words?

Could you trace the scaffolding thread — RLHF to Constitutional AI to vision to computer use to hybrid reasoning to twin-model safety redirects — without looking at the table?

⚠️ What's Missing or Uncertain

Every release after Claude 2 in this article is covered at bridge depth, not full case-study depth — consistent with the "consolidated comparison, skipping individual case studies" scope of this piece. Undisclosed cells reflect genuine gaps in Anthropic's own public materials. Several exact dates and figures for the most recent releases (Sonnet 5, Opus 5, Fable 5.1/Mythos 5.1) come from reporting very close to this article's writing date and should be checked against Anthropic's own current documentation for anything time-sensitive, such as pricing.

🔗 Reference Links

🎥 Recommended Videos

But What Is a Neural Network?

3Blue1Brown — the architecture family every Claude model, undisclosed details and all, is built on

Attention Is All You Need — Paper Explained

Yannic Kilcher — the 2017 Transformer paper underlying every model in both lineages this site has documented

🧭 Closing — Opacity as a Starting Position, Not an Ending One

🎯 The Bottom Line

Anthropic's fourteen-release, three-and-a-half-year lineage never had an open era to lose — its entire story had to be told through capability, scaffolding, and safety architecture instead of a shrinking disclosure trail. Constitutional AI, the ASL framework's first real activation a year ahead of GPT-6 Astra's comparable milestone, and the genuinely novel Fable/Mythos twin-model safety redirect are the through-line here — proof that a lab can build an entirely legible public safety narrative around a model it has never technically described in detail, which is precisely the opposite lesson the GPT lineage teaches.