Overview
Claude Fable 5.2 is Anthropic's flagship model released on November 9, 2026, built on the same underlying weights as Claude Mythos 5.2 unveiled the same day. What separates them is access, not capability: Fable ships with the full cybersecurity and biosecurity guardrails and is available on every paid plan including Pro, Max, Team and Enterprise, while Mythos lifts some dual-use restrictions and goes only to cybersecurity and life-science organizations cleared through Verification Programs.
Two months separate this release from 5.1 in early September. Pricing fell from $10 and $50 per million tokens to $8 and $40, a twenty percent cut, and cache reads dropped to $0.20. There is a footnote that is easy to miss: the new model burns noticeably more tokens on long tasks, so on a per-task basis the cost lands roughly where 5.1 did. The unit price fell; the invoice did not.
The capability claims center on long-horizon work. It tops the Artificial Analysis Intelligence Index v5 at 65, ahead of Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. It scores 93.8% on ARC-AGI-2, 98.5% on ARC-AGI-1, 69.9% on FrontierSWE v2, 89.3% on OSWorld 2.1 and 74.2% on Humanity's Last Exam with tools. Arena's Code Arena WebDev board puts it first at 1832, a shade above Opus 5.5 at 1818, and Cognition's early FrontierCode 2.0 run scores it 49.1, also first.
The demonstrations Anthropic leaned on hardest are the ones that ran for a long time. A thousand Fable 5.2 agents working together formalized the full proof of Fermat's Last Theorem in 30 hours, while Opus 5.5 under the same setup reached 26% at 18 hours and then barely moved; the run emitted a third of the tokens of the September formalization and produced a proof one seventh as long. A second project with Diode Computers started from a single prompt and ended with a speakerphone, a voice assistant hub and a handheld digital pet, complete with four-layer PCB routing, CAD enclosures and firmware, all manufactured straight from the design files and still working after weeks of use.
Anthropic's own guidance is restrained: start most coding and knowledge work at Opus 5.5 with medium effort, and reserve Fable 5.2 for open-ended technical work where the path is not clear. That sentence is effectively the product positioning. Twice the unit price, and a willingness to keep digging where no answer exists yet.
Key Features
- Long-horizon coding and terminal work: FrontierSWE v2 at 69.9%, OSWorld 2.1 at 89.3%, Humanity's Last Exam with tools at 74.2%. Cognition's FrontierCode 2.0 scores it 49.1, first on that benchmark.
- Top of Code Arena WebDev: 1832 at the highest reasoning effort, against 1818 for Opus 5.5 and 1768 for Gemini 4 Carbon. Cursor's own CursorBench run gives 56.9% at default effort and 60% at maximum.
- Three ARC-AGI results: ARC-AGI-2 at 93.8% costing $0.98 per task, ARC-AGI-1 at 98.5% costing $0.36 per task, and ARC-AGI-3 at 81.2% under the standard harness.
- Multi-agent teams on long tasks: Anthropic ran a thousand agents for 30 hours to formalize Fermat's Last Theorem. Google Cloud positions it for multi-repo work and unattended runs lasting hours to days with subagent coordination.
- 1M context and faster output: A one-million-token window with output up to 30% faster than 5.1. A separate fast mode doubles both output speed and price; set speed to fast in the API, or toggle it in the model picker inside Claude Code.
- Preserved thinking on every account: To hinder distillation, all API accounts now retain thinking traces by default. Editing earlier context may return a 400, so Anthropic points to compaction, context editing or mid-conversation system messages instead; Claude Code and the Agent SDK handle it automatically.
- Hardware and scientific computing: PCB Gym scores 69%, and all three Diode devices went straight from design files to manufacturing. Anthropic also says the model settled two open problems in black hole theory, one of them a calculation that had stood for 23 years.
Use Cases
- Large-scale refactors across several repositories or one oversized monorepo, where the model has to work out dependencies itself.
- Unattended runs lasting hours to days, including subagent division of labor and cross-review.
- Formal proofs and scientific computation, translating natural-language arguments into a language a checker can verify step by step.
- Hardware prototyping, from spec and schematic through board layout to enclosure and firmware in one pass.
- Reviewing codebases or document sets at the million-token scale.
Pros
- Long-horizon completion improved sharply: the Fermat demo hit 100% in 30 hours while the comparison model stalled at 27%.
- Unit price down twenty percent and cache reads at $0.20 per million tokens, which lightens long sessions.
- ARC-AGI-2 at 93.8% sits third in the top tier, behind GPT-6 Astra at 95.0% and GPT-6.1 Sol at 94.2%.
- A 1M window combined with a 30% output speedup cuts waiting time on big jobs.
- Major clouds and agent tools shipped support the same day, with Cursor, Devin, Perplexity and OpenRouter all callable immediately.
- A fast mode lets you trade money for latency when a deadline matters.
Pricing
Input costs $8 per million tokens and output $40, with cache reads at $0.20, twenty percent below the $10 and $50 of 5.1. Fast mode doubles output speed and doubles the price, billed against extra usage credits inside Claude Code. Per task, the heavier token consumption means cost lands roughly level with 5.1, so the cut shows up in the unit price rather than the invoice. Pro, Max, Team and Enterprise plans all include access, and Anthropic says the longer you have subscribed, the more Fable resets you receive.
Summary
Claude Fable 5.2 is Anthropic's flagship refresh ahead of its IPO, sharing weights with Mythos 5.2 and differing only in access policy. It arrives with reproducible demonstrations in long-horizon coding, formal proof work and hardware design, tops the AA Intelligence Index at 65, and cuts unit pricing by twenty percent while holding per-task cost flat. It suits multi-repo refactors, unattended long runs and open-ended technical problems; routine coding still starts at Opus 5.5.
Version History
- Claude Fable 5.2 ships alongside Mythos 5.2 (2026-11-09): Anthropic released Claude Fable 5.2 on November 9, 2026, sharing underlying weights with Mythos 5.2. Fable carries the full cybersecurity and biosecurity guardrails and opens to every paid plan. Pricing fell from $10/$50 to $8/$40 per million tokens with cache reads at $0.20, output is up to 30% faster, and context reaches 1M. Artificial Analysis Intelligence Index v5 places it first at 65 (Gemini 4 Carbon 60, Opus 5.5 59), with ARC-AGI-2 at 93.8%, ARC-AGI-1 at 98.5%, FrontierSWE v2 at 69.9%, OSWorld 2.1 at 89.3%, HLE with tools at 74.2% and Code Arena WebDev at 1832 in first place. Official demos include a thousand agents formalizing Fermat's Last Theorem in 30 hours and three manufacturable devices designed with Diode from prompts. The release also adds a fast mode at 2.5x speed and 2x price, and extends preserved thinking to all API accounts.