1. 🧭 The Second Lineage, A Different Founding Bet
Anthropic was founded in 2021 explicitly as a safety-focused split from OpenAI. That's a founding stance, not a guarantee of a different technical roadmap — the question worth actually checking, using the same method as our GPT series post, is whether that stance shows up as a genuinely different sequence of levers pulled, or whether both labs converged on the same moves at roughly the same times regardless of stated philosophy.
2023 → 2026
Claude 1 to Claude Fable 5.1 — a three-year lineage, shorter than GPT's but covering the same major eras
Feb 2025
Claude 3.7 Sonnet's "extended thinking" pivot — Anthropic's own test-time-compute moment, months after OpenAI's o1
14h30m
Claude Opus 4.6's METR autonomous time-horizon record, Feb 2026 — briefly the industry's longest
Dec 2022
Constitutional AI published — Anthropic's own post-training lever, distinct from RLHF, present since before Claude 1 even shipped
2. 🧱 Quick Reference
Same 9-layer stack and 6 levers as the rest of this series — full definitions in The AGI Threshold and The Six Levers.
Lever 1: Pre-training Scale
Lever 2: Compute/Efficiency
Lever 4: Post-Training (incl. Constitutional AI)
Lever 3: Test-Time Compute
Lever 5: Agentic Scaffolding
Lever 6: Memory/Safe Autonomy Research
3. 📅 The Full Claude Timeline
Founding/Scale
Tiering/Efficiency
Post-Training/Agentic
Reasoning Era
Safety-Architecture Era
FoundingMar 14, 2023
Claude 1 / Claude Instant
First public API launch — pre-training scale, plus Constitutional AI from day one
LeverStandard pre-training scale, but already layered with Constitutional AI (published Dec 2022) — Anthropic's own post-training alternative to RLHF, present from the very first release rather than added later as a pivot.
FoundingJul 11, 2023
Claude 2
First model available to the general public, alongside the claude.ai chat interface
LeverScale-driven capability increase, but the more consequential move was product access — the public chat interface — a signal Anthropic was already thinking about consumer-facing scaffolding, not just API-only research releases.
TieringMar 4, 2024
Claude 3 Family (Haiku, Sonnet, Opus)
Three tiers, one launch — cost/capability tradeoff built into the product line itself
LeverA structural efficiency-lever bet made explicit from the start, unlike OpenAI's later bolt-on "Turbo/mini" tiers — Haiku, Sonnet, and Opus launched together as one family, with Haiku reaching general availability March 13, 2024.
Post-Training/AgenticJun 20 – Oct 22, 2024
Claude 3.5 Sonnet (v1 → v2)
A mid-tier model beating the prior flagship, then "computer use" arrives
LeverThe clearest post-training/efficiency story in the lineage — a Sonnet-tier model outperforming the prior Opus-tier flagship. The v2 update (Oct 22, 2024) shipped "computer use" in public beta — Claude directing a cursor and keyboard by reading the screen — the first publicly available model with this capability, a direct Lever 5 (scaffolding/tool use) move.
Reasoning EraFeb 25, 2025
Claude 3.7 Sonnet
The first hybrid reasoning model — near-instant or extended thinking, visible to the user
LeverAnthropic's own test-time-compute pivot, arriving roughly five months after OpenAI's o1 preview — a "hybrid" design letting the same model choose near-instant or extended, step-by-step visible thinking, with fine-grained API control over thinking duration. Shipped alongside Claude Code.
Safety-Architecture EraMay – Nov 2025
Claude 4 (Opus 4/Sonnet 4) → Sonnet 4.5 → Opus 4.5
Agentic coding as the center of gravity, capped by Opus 4.5's "best in the world" claim
LeverClaude 4 family launched May 22, 2025; Sonnet 4.5 in September 2025; Haiku 4.5 in October 2025; Opus 4.5 on November 24, 2025, positioned explicitly as "the best model in the world for coding, agents, and computer use" — three consecutive releases doubling down on Lever 5, not a new pivot.
Safety-Architecture EraFeb – Jul 2026
Claude Opus 4.6 → Opus 4.8
The time-horizon record, Dynamic Workflows, and Fast mode
LeverOpus 4.6 (Feb 2026) set the METR autonomous time-horizon record at 14h30m — the clearest Lever 5/3 combination result in the lineage. Opus 4.8 added "Dynamic Workflows" (hundreds of parallel subagents for codebase-scale migrations) and a 2.5x-faster, one-third-cost "Fast mode" — efficiency (Lever 2) applied directly to agentic throughput.
Current FrontierJul – Sep 2026
Claude Opus 5 → Claude Fable 5.1
Cheaper frontier performance, then "adaptive thinking that stays on"
LeverOpus 5 (late July 2026, $5/$25 per million tokens) delivered near-Fable performance at roughly half the price — a pure Lever 2/4 efficiency story. Fable 5.1 (Sept 1, 2026) added a 1M-token context window and "adaptive thinking that stays on," explicitly positioned for autonomous multi-day work — and Anthropic stated Opus 5 has no more concerning alignment properties than Fable 5, a rare direct Lever 6 safety claim in a release note.
4. 🚀 Claude 1 & Claude 2: Scale, Plus Constitutional AI From Day One (2023)
Claude 1 / Claude Instant → Claude 2
Mar 14, 2023 → Jul 11, 2023
Lever 1: Pre-training ScaleLever 4: Constitutional AI
The single most notable difference from OpenAI's founding sequence: Anthropic's post-training lever (Constitutional AI, published December 2022 — training a model against a written "constitution" rather than solely human preference labels) was already in place before Claude 1 ever shipped, not added as a later pivot the way OpenAI's RLHF/InstructGPT moment came after two pure-scale releases (GPT-1/2/3). Claude 2 (July 2023) paired continued scale with the first public claude.ai chat interface.
Stack layers touched: Reasoning, Generalization, early Reliability (via constitutional training)
5. 🎭 The Claude 3 Family: Tiering as Strategy, Not Afterthought (2024)
Claude 3: Haiku, Sonnet, Opus
March 4, 2024
Lever 2: Compute/Efficiency
Three tiers launched together, not staggered — a structural bet that cost/capability tradeoffs should be a first-class product decision, not a later cost-optimization pass. This is a genuinely different sequencing choice than OpenAI's, whose efficiency-tier releases (Turbo, 4o, o3-mini) each arrived well after an initial single flagship model, as retrofits rather than launch strategy.
Stack layers touched: Reasoning, Multimodality (image input across all three tiers)
6. 🧵 Claude 3.5 Sonnet: Punching Above Its Tier (2024)
Claude 3.5 Sonnet (v1 → v2)
Jun 20, 2024 → Oct 22, 2024
Lever 4: Post-TrainingLever 5: Computer Use
A mid-tier (Sonnet) model matching or beating the prior flagship (Opus) tier is the cleanest available evidence that post-training refinement, not base-model scale, drove the jump — the same InstructGPT-style pattern seen in the GPT lineage, here happening one tier down rather than within the same tier. The v2 update's public beta "computer use" feature (Claude reading a screen and directing a cursor/keyboard) was the first such capability shipped by any lab to the public, making Anthropic first-to-market on this specific Lever 5 capability, ahead of OpenAI's equivalent agentic push.
Stack layers touched: Tool & Computer Use (new, first-to-market), Generalization
Claude 3.5 Sonnet beating the prior Opus flagship is Anthropic's InstructGPT moment — proof a smarter harness on an existing base model can matter more than the next scale-up.
7. 🤔 Claude 3.7 Sonnet: The Reasoning Pivot (Feb 2025)
Claude 3.7 Sonnet
February 25, 2025
Lever 3: Test-Time Compute
Marketed as the first hybrid reasoning model — a single model that can respond near-instantly or think at length, visible to the user, rather than requiring a separate reasoning-specific model line the way OpenAI split o1/o3 from GPT-4/4o. This is a real design difference in how the same lever got implemented: Anthropic chose one adjustable model over two separate product lines. Shipped alongside Claude Code, tying the reasoning pivot directly to agentic coding from the outset.
Stack layers touched: Reasoning (major), Reliability (partial, via visible extended thinking)
8. 🏗️ Claude 4, Sonnet 4.5 & Opus 4.5: Agentic Coding as the Center of Gravity (2025)
Claude 4 Family → Sonnet 4.5 → Haiku 4.5 → Opus 4.5
May 22 – Nov 24, 2025
Lever 5: Agentic ScaffoldingLever 3: Test-Time Compute
Four releases in six months, each reinforcing the same lever combination rather than pivoting to a new one — a sustained push, not a scale-then-plateau-then-pivot pattern. Opus 4.5 (November 24, 2025) was positioned as "the best model in the world for coding, agents, and computer use," explicit language that agentic capability, not raw benchmark reasoning, was the primary competitive claim being made.
Stack layers touched: Tool & Computer Use, Long-Horizon Planning, Reasoning (refined)
9. 🛡️ Opus 4.6 → 4.8: The Time-Horizon Record and Safety Architecture (2026)
Claude Opus 4.6 → Opus 4.8
Feb 2026 → mid-2026
Lever 5Lever 2
Opus 4.6 briefly held the METR 50%-reliability autonomous time-horizon record at 14 hours 30 minutes — the clearest quantified Lever 5/3 combination result in this entire lineage, until later frontier releases (including GPT-6 Astra's 24-hour Portal completion) surpassed it. Opus 4.8 introduced "Dynamic Workflows" — hundreds of parallel subagents coordinating on codebase-scale migrations — and a "Fast mode" running 2.5x faster at one-third the cost, an efficiency lever applied specifically to agentic throughput rather than to the base model's raw benchmark scores.
Stack layers touched: Long-Horizon Planning (record-setting), Tool & Computer Use, efficiency across all layers
10. 🟣 Opus 5 & Claude Fable 5.1: Cheaper Frontier, Adaptive Thinking (2026)
Claude Opus 5 → Claude Fable 5.1
Late Jul 2026 → Sep 1, 2026
Lever 2: EfficiencyLever 6: Safety Claim
Opus 5 shipped at $5/$25 per million tokens with near-Fable performance at roughly half the price — Frontier-Bench 43.3 vs. Opus 4.8's 18.9, ARC-AGI-3 at 30.2% vs. 1.5% — almost entirely a post-training/efficiency story rather than a base-model scale story. Fable 5.1 (September 1, 2026) added a 1M-token context window, 128K output, and "adaptive thinking that stays on," explicitly positioned for long-running agentic and research work. Notably, Anthropic stated Opus 5 shows no more concerning alignment properties than Fable 5 — a direct, publicly stated Lever 6 claim, and still recommended Fable 5.1 specifically for autonomous multi-day projects, treating "most capable" and "most trusted for long autonomy" as explicitly separate axes.
Stack layers touched: Reasoning, Generalization, Long-Horizon Planning, explicit Safe Autonomy statement
11. 📊 The Full Picture: Claude Releases × Stack Layers
| Release | Reasoning | Generaliz. | Multimodal | Memory | Tool Use | Long-Horizon | Reliability | Metacog. | Safe Auton. |
|---|---|---|---|---|---|---|---|---|---|
| Claude 1/2 | |||||||||
| Claude 3 Family | |||||||||
| Claude 3.5 Sonnet | 1st CU | ||||||||
| Claude 3.7 Sonnet | |||||||||
| Claude 4/4.5 | |||||||||
| Opus 4.6/4.8 | 14h30m | ||||||||
| Opus 5/Fable 5.1 | 30.2% | multi-day | stated claim |
🟢 Strong 🟡 Good-partial 🟠 Partial 🔴 Early/weak ⚪ Not yet present — editorial synthesis based on sourced release claims above
Compared to the GPT matrix in our prior post, the Safe Autonomy column here starts higher (Constitutional AI from Claude 1) and climbs more steadily rather than staying flat — the one column where Anthropic's lineage looks visibly different from OpenAI's, consistent with its founding stance actually showing up in the release history, not just the marketing.
12. 🔁 Anthropic vs. OpenAI: Same Levers, Different Sequencing
Scale + Constitutional AI
Claude 1/2
→
Claude 1/2
Tiering
Claude 3
→
Claude 3
Post-Training + Tools
3.5 Sonnet
→
3.5 Sonnet
Test-Time Compute
3.7 Sonnet
→
3.7 Sonnet
Sustained Agentic Push
4.x → 4.8
→
4.x → 4.8
Efficiency + Stated Safety
Opus 5/Fable 5.1
Opus 5/Fable 5.1
🟣 Where Anthropic Differs
Post-training safety lever (Constitutional AI) present from the first release, not added later after a pure-scale phase
Product tiering (Haiku/Sonnet/Opus) launched as a structural strategy from day one of Claude 3, not retrofitted the way OpenAI added Turbo/mini/o3-mini variants
Once the agentic-scaffolding lever was found (3.5 Sonnet's computer use), Anthropic sustained it across five consecutive releases rather than pivoting away, and paired it with public safety-architecture claims (permission-first computer use, real-time classifiers) more consistently than the other lab in this series
🔶 Where They Converge
Both hit a test-time-compute pivot within about five months of each other (OpenAI's o1 preview Sep 2024; Anthropic's 3.7 Sonnet Feb 2025) — suggesting the lever's arrival was industry-timed, not lab-specific insight
Both labs' most recent releases lean heavily on Lever 5 (scaffolding/agentic) as the primary axis of competition, not new pre-training scale
Neither lineage has a fully solved Metacognition column — Anthropic's is consistently higher than OpenAI's, but still capped at "good-partial" at best across every release examined here
13. 🗣️ "Powerful AI" as a Lever Choice, Not Just a Word Choice
Dario Amodei's preference for "powerful AI" over "AGI" (covered in our Astra, Claude & Gemini post) reads differently once placed against this release history. A lab whose own timeline shows a steadily, if incompletely, rising Safe Autonomy column — and which stated directly in a release note that Opus 5 has "no more concerning alignment properties" than Fable 5 — has more grounds than most to distinguish "broadly capable" from "safely autonomous" as genuinely separate claims. The terminology preference lines up with the actual lever history, rather than functioning as pure positioning.
14. 🔮 What's Next for Claude
Memory Is the Obvious Next Reach — Here Too
Same conclusion as the GPT post: this lineage has never seriously pulled the memory/continual-learning lever. Fable 5.1's "adaptive thinking that stays on" is adjacent but not the same as durable cross-session memory.
Formalizing the Safety Architecture
Given the consistent pattern of pairing capability releases with safety-architecture statements, expect Anthropic's next release to make an even more explicit, possibly third-party-audited safe-autonomy claim rather than a purely internal one.
Efficiency Keeps Compounding
Opus 5's half-price-at-near-parity result and Opus 4.8's Fast mode both point the same direction — expect continued aggressive cost reduction on agentic workloads specifically, since that's where the commercial pressure from OpenAI's GPT-5.6 tiering is most direct.
15. 🧭 Verdict
🎯 The Bottom Line
Three years of Claude releases show the same six levers as the GPT lineage, pulled in a genuinely different order and, in one respect, a different rhythm: Anthropic front-loaded its safety-post-training lever (Constitutional AI) instead of arriving at it after a pure-scale phase, built product tiering into its very first multi-model launch instead of retrofitting it, and — once it found the agentic-scaffolding lever with 3.5 Sonnet's computer use — sustained investment in it across five consecutive releases while consistently pairing capability claims with public safety-architecture statements. The founding stance shows up as sequencing, not just as marketing language — the Safe Autonomy column in this lineage is higher and steadier than in the GPT matrix, even though it never gets past "good-partial." Where the two labs converge — the test-time-compute pivot arriving within months of each other, both current flagships leaning hardest on scaffolding — is arguably the more important finding: some lever transitions look industry-timed rather than lab-specific, meaning the next one (memory and continual learning, on the evidence of both lineages examined so far) is likely coming for both labs at roughly the same time too.