Claude Fable 5

Overview

Claude Fable 5 is Anthropic's first Mythos-class model made generally available. Released on June 9, 2026, it uses the same underlying model as Claude Mythos 5, but adds safeguards that let Anthropic expose frontier cybersecurity, scientific, coding, vision, and research capabilities to a broad audience.

This distinction is the key to understanding the product. Mythos 5 is the trusted-access deployment for selected cyberdefenders and life-science researchers. Fable 5 is the general API model: when its classifiers detect certain cybersecurity, biology and chemistry, or model-distillation requests, the request is handled by Claude Opus 4.8 instead. For the more than 95% of sessions that do not trigger fallback, Anthropic says Fable performs effectively the same as Mythos 5.

Key capabilities

Prompt caching uses 0.1x cache reads, 1.25x 5-minute writes, and 2x 1-hour writes relative to base input. Mythos-class traffic is retained for 30 days for safety monitoring and is not used to train new Claude models. See live pricing in the model catalog.

Long-horizon software engineering

Anthropic positions Fable 5 around autonomous work that lasts much longer than a normal chat turn. Stripe reported that it completed a codebase-wide migration across a 50-million-line Ruby repository in one day, compressing work that would otherwise have taken a full team more than two months.

The important change is not only a higher success rate. On Cognition's FrontierCode evaluation, Fable 5 produced the strongest result among frontier models even at medium effort, while using tokens more efficiently than earlier Claude models. Early users similarly describe fewer turns, stronger self-validation at high effort, and better generalization to unfamiliar tools.

Knowledge work, vision, and memory

Fable 5 leads Anthropic's comparisons on senior-level financial reasoning, including document analysis, chart and table interpretation, root-cause analysis, and expected-value reasoning. One early customer reported the first result above 90% on its long-running analytics benchmark, a 10-point gain over Opus; another found spreadsheet runs finished 25%-30% faster than Opus 4.8 with fewer turns.

Vision is not treated as a side feature. Fable 5 can extract exact values from dense scientific figures and rebuild a web application's source code from screenshots. Anthropic's clearest demonstration is Pokémon FireRed: earlier Claude models needed a substantial helper harness, while Fable 5 completed the game using raw screenshots and a minimal vision-only loop.

Long-context work also benefits from persistent memory. In Slay the Spire, giving Fable access to file-based notes improved its performance three times more than the same technique helped Opus 4.8, and Fable reached the final act three times more often. This points to a model that can accumulate experience across a run rather than repeatedly solving each step from scratch.

Scientific capability and the Fable boundary

The shared Fable/Mythos model is unusually capable in life sciences. Anthropic reports that Mythos 5 accelerated parts of protein-design work by roughly 10x; 9 of 14 protein targets yielded strong drug-design candidates. In blinded comparisons, Anthropic scientists preferred its molecular-biology hypotheses over Opus-class models about 80% of the time.

In a separate week-long genomics project, Mythos assembled single-cell data covering millions of cells across 138 animal species, then trained a model that outperformed a recent Science-published system despite being 100x smaller. These results explain both Fable 5's research potential and its restrictions: general API requests in biology and chemistry currently fall back broadly to Opus 4.8, while less-restricted Mythos access is reserved for approved researchers.

Safeguards and integration behavior

Fable 5's classifiers cover offensive cybersecurity, risky biology and chemistry, jailbreak attempts, and suspected capability distillation. They are deliberately conservative at launch, so harmless requests can occasionally trigger fallback. Anthropic says fallback occurs in fewer than 5% of sessions on average.

When this happens, the user receives an answer from Opus 4.8 rather than a refusal from Fable. Applications should therefore surface the actual responding model and avoid assuming that every successful Fable request was completed by Fable itself. Anthropic reports that an external bug bounty found no universal jailbreak after more than 1,000 hours of testing, while acknowledging that the classifiers will continue to be refined to reduce false positives.

When to use it

• Ambitious coding agents: large migrations, unfamiliar tools, multi-agent workflows, and long autonomous validation loops.
• Complex knowledge work: financial analysis, legal review, research synthesis, and spreadsheet-heavy workflows.
• Vision-first tasks: scientific figures, screenshot-to-code reconstruction, and computer-use agents with minimal scaffolding.
• Persistent long-running agents: workflows that learn from notes, files, and previous attempts.
• Safety-aware products: integrations that can identify Opus fallback and accept the 30-day retention policy.

Choose Fable 5 when task length, autonomy, and cross-domain reasoning matter enough to justify the top Claude tier. Choose Opus 4.8 when lower cost, simpler deployment behavior, or avoiding Mythos-class retention and fallback rules matters more.

CrossModel exposes Claude Fable 5 through Anthropic-compatible /v1/messages and OpenAI-compatible /v1/chat/completions. Current pricing is available in the model catalog.