Overview
Claude Fable 5 is Anthropic's June 9, 2026 flagship and its most capable widely released model to date — built for the most demanding reasoning and long-horizon agentic work. Fable 5 is a Mythos-class model made safe for general public use: it carries the raw capability of the restricted Claude Mythos 5 but adds safety classifiers that route queries on a narrow band of sensitive topics to Anthropic's next-most-capable model, Claude Opus 4.8. In Anthropic's own measurement, roughly 95% of sessions never trigger that fallback at all.
The model is generally available on the day of release across claude.ai, Claude Code, the Claude API as claude-fable-5, the Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Fable 5 ships a 1M-token context window by default and supports up to 128k output tokens per request. Two integration behaviors set it apart from Opus, Sonnet, and Haiku: adaptive thinking is always on (you cannot disable thinking), and the raw chain of thought is never returned — you receive summarized thinking only if you request it.
What's New
Mythos-Class Capability, Generally Available
Fable 5's headline is access, not just benchmarks. It shares the underlying model with Mythos 5, the limited-release variant offered through Project Glasswing, but wraps it in safety classifiers so the same frontier capability can ship to every developer. When Fable 5 declines a request, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 — not an error — and reports which classifier declined it. A beta fallbacks parameter (and SDK middleware) can automatically retry the request on Claude Opus 4.8, with fallback credit helping avoid paying the prompt-cache write cost twice.
Long-Horizon Agentic Work
Fable 5 is tuned for tasks that run for hours and span millions of tokens. During early testing, Stripe reported that Fable 5 compressed months of engineering into days — including a codebase-wide migration of a 50-million-line Ruby codebase completed in a single day, work that would otherwise take a team more than two months by hand. It posts the highest score among frontier models on Cognition's FrontierCode evaluation, and pairs the long context with strong self-correction across extended tool-use loops.
Long-Context Memory and Reasoning
The 1M-token context is not just large — Fable 5 stays focused inside it. In Anthropic's Slay the Spire testing, giving Fable 5 access to persistent file-based memory improved its performance roughly three times more than the same memory helped Opus 4.8, and Fable 5 reached the game's final act three times more often — a proxy for holding and acting on details introduced far earlier in a long session.
State-of-the-Art Vision
Fable 5 sets a new bar on complex vision tasks. In one benchmark of agentic perception, it completed Pokémon FireRed using vision alone — navigating, reading the screen, and acting without any game-state API. For practical work, that translates to stronger reading of dense diagrams, screenshots, and document layouts inside agentic loops.
Knowledge Work and Scientific Research
Beyond code and vision, Fable 5 records the highest score on Hebbia's Finance Benchmark for senior-level financial reasoning — positioning it for analyst-grade knowledge work rather than general chat alone. Anthropic credits the strongest molecular-biology hypothesis generation to the Mythos-class capability; because Fable 5's safeguards can route some biology and chemistry prompts to Opus 4.8, the most autonomous frontier research sits with Claude Mythos 5.
Performance Benchmarks
Headline numbers from Anthropic's release:
| Benchmark / Task | Claude Fable 5 | Baseline |
|---|---|---|
| FrontierCode (Cognition) | Highest among frontier models | — |
| Slay the Spire + persistent memory | Memory helped Fable 5 ~3× more; reached final act 3× more often | vs. Opus 4.8 |
| 50M-line Ruby migration (Stripe) | ~1 day | More than two months for a team by hand |
| Hebbia Finance Benchmark | Highest score | Senior-level reasoning |
| Vision (Pokémon FireRed) | Completed via vision alone | — |
| Safeguard fallback rate | <5% of sessions | ~95% no fallback |
The coding and long-context numbers matter most for teams running autonomous AI agent pipelines or AI code checker workflows, where context drift and silent failures previously capped how long an agent could run unattended.
Migration Guide
Anthropic publishes a dedicated path for moving from Opus 4.8 to Fable 5. The capability ceiling rises, but two behavior changes warrant attention before cutover:
Thinking Behavior
Adaptive thinking is the only thinking mode on Fable 5 — thinking: {"type": "disabled"} is not supported, and you control depth through the effort parameter instead. The raw chain of thought is never returned (thinking.display defaults to "omitted"); set display: "summarized" if you need readable summarized thinking. Pass thinking blocks back unchanged in multi-turn conversations on the same model.
Refusals, Fallback, and Data Retention
Because Fable 5 includes safety classifiers, design for the refusal path: handle stop_reason: "refusal" as a 200 response and wire up the beta fallbacks parameter or SDK middleware so refused requests retry automatically — at launch, the permitted fallback target for Fable 5 is Claude Opus 4.8. You are not billed for a request refused before any output is generated. Note also that Fable 5 is a Covered Model — it carries 30-day data retention and is not available under zero data retention.
Pricing & Plans
| Plan | Price |
|---|---|
| API – Input | $10 per million tokens |
| API – Output | $50 per million tokens |
| Claude.ai Pro | Included within plan limits |
| Claude.ai Max | Included with higher usage allowances |
| Team / Enterprise | Contact Anthropic for volume pricing |
| Cloud availability | Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry |
Cost note: Fable 5 sits above the $5/$25 of the current Opus-tier model, Claude Opus 4.8, and the $3/$15 of Sonnet 4.6. Reserve it for the hardest reasoning and longest-running agentic workloads, and route routine traffic to a cheaper tier — the
fallbacksmechanism makes that routing straightforward.
Best For
- Engineering teams running autonomous, multi-hour coding agents over large codebases where Opus-class models hit context or reliability ceilings
- AI agent platforms that need a model to stay coherent across millions of tokens and long tool-use loops
- Developers using Claude Code on migrations, refactors, and codebase-wide changes that previously took weeks
- Finance, research, and knowledge-work teams needing senior-level reasoning over dense documents and data
- Vision-heavy agentic workflows — browser agents, screen automation, diagram analysis — that depend on reading the screen accurately
- Teams that can route only their hardest traffic to Fable 5 and send routine work to cheaper models — Claude Opus 4.8 (the launch fallback target), Opus 4.7, or Sonnet 4.6
FAQ
What is Claude Fable 5?
Claude Fable 5 is Anthropic's most capable widely released model, launched June 9, 2026 for demanding reasoning and long-horizon agentic work. It is a Mythos-class model made safe for general use: it carries the same underlying capability as the restricted Claude Mythos 5 but adds safety classifiers that fall back to Opus 4.8 on sensitive prompts in under 5% of sessions.
How is Fable 5 different from Claude Mythos 5?
They share the same underlying model. Fable 5 is generally available with safety classifiers and is the model most teams should use. Mythos 5 has its safeguards lifted in some areas and is offered only in limited availability through Project Glasswing — initially to approved cyberdefenders and infrastructure providers, with a trusted biology access program planned. If you do not have Glasswing access, Fable 5 is the generally available Mythos-class model.
How much does Claude Fable 5 cost?
Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — roughly double the input price of Opus 4.7. On claude.ai it is included within Pro, Max, Team, and Enterprise plan limits.
What context window does Fable 5 support?
Fable 5 supports a 1M-token context window by default and up to 128k output tokens per request. In Anthropic's Slay the Spire memory testing, persistent file-based memory improved Fable 5's performance roughly three times more than it improved Opus 4.8.
Why did Fable 5 return a refusal instead of an answer?
Fable 5's safety classifiers can decline certain requests. When that happens the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response (not an error) and reports which classifier declined it. You can pass the beta fallbacks parameter or use SDK middleware to automatically retry the request — at launch the permitted fallback target is Claude Opus 4.8 — and you are not billed for a request refused before any output is generated.
Where can I use Claude Fable 5?
Fable 5 is generally available on claude.ai, Claude Code, the Claude API as claude-fable-5, the Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
