Anthropic says Claude Fable 5.2 and Claude Mythos 5.2 use the same underlying model but differ in safeguards and access. Fable retains protections against high-risk cyber activity and most research-biology tasks; vetted organizations can apply for Mythos through verification programs. Anthropic’s announcement also positions Fable for demanding, open-ended technical work, not as the default for every coding task.

That launch is a separate development from NeoTeo’s earlier coverage of the Mythos 5 federal-policy dispute.

Choose Fable for the work that warrants its higher-effort approach; for routine coding and knowledge tasks, Anthropic recommends starting with Opus 5.5.

Fable and Mythos share a model, not an access policy

The distinction Anthropic describes is chiefly about safeguards and eligibility. Fable blocks high-risk cyber activity and most research-biology tasks. Mythos has certain cyber and biology safeguards removed, and access is limited to vetted organizations applying through Anthropic’s verification programs.

OfferingUnderlying modelSafeguards Anthropic describesAccess Anthropic describes
Claude Fable 5.2Shared with Mythos 5.2Blocks high-risk cyber activity and most research-biology tasksAnthropic says it is available across its platforms and cloud partners
Claude Mythos 5.2Shared with Fable 5.2Certain cyber and biology safeguards are absentVetted organizations can apply through verification programs

Fewer safety-system blocks do not mean Fable’s safeguards have disappeared. Anthropic says early testing across claude.ai, Claude Code, and Cowork found about 55% fewer unnecessary blocks than with Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology. The company says Fable still blocks high-risk cyber activity and most research-biology tasks.

Anthropic recommends Opus 5.5 as the starting point for most coding and knowledge work. It presents Fable for more ambitious, open-ended technical tasks, including as an adviser or orchestrator for other models. That distinction makes the choice less about picking one winner and more about matching the model to the job.

What the coding scores say—and what they don’t

On Anthropic’s FrontierCode 2.0 evaluation, Fable 5.2 scored 45.3% at High effort; Opus 5.5’s best result was 45.5% at Max. Anthropic says Fable’s task cost was roughly half in that comparison. Because the models were tested at different effort settings, those figures describe that benchmark and setup, not a universal price or performance advantage. Fable scored 49.1% at Xhigh on the same evaluation.

Anthropic also reports an 83.5 score for Fable 5.2 on BenchCAD 2.1 with tools at Max effort, compared with 65.3 for Opus 5.5 and 44.4 for Fable 5.1. The benchmark measures how closely hardware designs match reference geometry and circuits, on a scale from 0 to 100 with partial credit. It is a different test from FrontierCode 2.0, so the scores should be read within their own benchmarks.

Anthropic lists standard rates of $8 per million input tokens and $40 per million output tokens. Its optional fast mode is listed at $16 per million input tokens and $80 per million output tokens, with generation speed of up to 2.5 times the standard mode. Anthropic lists fast mode for Amazon Web Services, Google Cloud, Microsoft Azure, and Microsoft Foundry. These are token rates, not subscription prices.

During launch week, Anthropic said eligible plans and seats received one to three Fable resets, depending on subscription tenure. For Pro plans and standard Team seats, Fable use counted for up to 50% of weekly limits during that promotion.

The science examples are specific claims, not broad discoveries

Anthropic says Fable 5.2 formalized an existing proof of Fermat’s Last Theorem in Lean, a language used to express proofs in a form a computer can check. The company says about 1,000 agents completed the formalization in 30 hours. That work turned an existing proof into a machine-checkable one; it did not discover the theorem.

Anthropic also attributes two theoretical-physics results to Fable 5.2. One calculation asks whether an abrupt black-hole phase transition also occurs in very small universes. The other identifies an exact limit for a technique used to rule out candidate quantum-gravity theories.

For a separate demonstration, Anthropic says Fable coordinated more than 1,000 subagents over eight days to build a prototype operating system that boots on a laptop. The company’s examples describe distinct tasks: formalizing a known proof, investigating narrowly framed physics questions, and coordinating work on a software prototype.