Claude Fable 5.1 API Guide: Pricing, Model ID, Benchmarks & Migration

Claude Fable 5.1 is Anthropic's highest-capability public model for long-horizon agents, difficult coding work, research, and document-heavy workflows. It was released on September 1, 2026. Anthropic describes Fable 5.1 and the restricted-access Mythos 5.1 as the same underlying model with different safeguards.

The API model ID is "claude-fable-5-1". It accepts text and images, returns text, supports a 1-million-token context window, and can generate up to 128,000 output tokens. Its training data cutoff is June 2026.

Fable 5.1 is not the default choice for every application. Anthropic recommends starting most workloads with Opus 5 and moving to Fable 5.1 when difficult reasoning or long-running agent work justifies the higher price and latency.

Fable 5.1 is generally available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. It is also available in eligible Claude Pro, Max, Team, and Enterprise plans, but a Claude subscription does not include API credits.

For the direct Anthropic API, sign in to the Anthropic Console, add billing if required, open API Keys, and create a key. Copy it immediately and store it as ANTHROPIC_API_KEY in an environment variable or secret manager. Never paste a production key into browser code, a public repository, screenshots, or support messages.

If you need a more detailed account walkthrough, use the Claude API key setup guide. The same Anthropic key can call multiple Claude models when your account has access; there is no separate key format for Fable 5.1.

Standard Fable 5.1 pricing is $10 per million input tokens and $50 per million output tokens. A 5-minute cache write costs $12.50 per million tokens, a 1-hour write costs $20, and a cache read costs $0.25. Batch processing reduces standard input and output rates by 50%.

For example, a request with 1 million uncached input tokens and 100,000 output tokens costs $15 before tool charges: $10 for input plus $5 for output. The same token volume through Batch costs $7.50. Reading a 1-million-token cached prefix costs $0.25, although the original cache write is charged separately.

These examples are arithmetic based on list prices, not a promise about the cost of a finished task. Adaptive thinking, retries, tools, cache creation, partner-platform fees, and regional routing can change the final bill. US-only inference applies a 1.1× multiplier.

Install Anthropic's official SDK, set ANTHROPIC_API_KEY outside your source code, and use the exact model ID. The minimal examples below leave tool use and advanced output controls out so you can verify account access first.

Fable 5.1 uses adaptive thinking on every request. Do not copy a manual thinking budget from older examples; requests that try to disable adaptive thinking or set an unsupported manual budget return a 400 error. Start with the default high effort, then test other supported effort settings only when you have measured quality and cost on your own tasks.

Opus 5 costs $5 per million input tokens and $25 per million output tokens, half the standard Fable 5.1 rates. It also has moderate rather than slow relative latency. This makes Opus 5 the practical starting point for most production systems.

Move to Fable 5.1 when the task is difficult to recover from, requires many coordinated steps, or repeatedly fails with Opus 5 at high effort. Useful candidates include autonomous codebase work, long research runs, large document transformations, and computer-use workflows. Run an evaluation on completed-task quality and total cost instead of comparing one response in isolation.

For a broader lineup comparison, see Claude models and API pricing.

Anthropic reports 55.8% on Terminal-Bench 4.0, 31.4% on AutomationBench, 73.4% on CursorBench 3.2, and 65.0% on Humanity's Last Exam with tools. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6%, compared with 24.7% for Fable 5 and 29.0% for Opus 5 in Anthropic's table.

Treat these as first-party results, not independent proof that Fable 5.1 is best for every workload. Anthropic notes that production safeguards were enabled and that some interventions produced zero scores or fallback responses. The OSWorld task release also differs from older evaluations, so those results are not directly comparable across every report.

The useful signal is narrower: Anthropic's tests show a large improvement on long-horizon terminal, automation, and agentic coding tasks. Before switching a production system, test the prompts, tools, failure modes, and budget that matter to your application.

The first code change is simple: replace "claude-fable-5" with "claude-fable-5-1". Managed Agents generally need only that model-name update. Messages API integrations must also account for stricter conversation-history and tool rules.

Fable 5.1 does not support forced tool choice with tool_choice set to "any" or a named tool; those requests return 400. Use "auto" or "none", then enforce argument shape with strict tool use or structured outputs where appropriate. Assistant-message prefilling is also unsupported.

Keep conversation history append-only. Earlier Claude models cannot read Fable 5.1 thinking blocks, and editing or reordering an earlier system message, tool definition, or turn can invalidate saved thinking blocks. Preserve those blocks exactly when continuing a conversation with Fable 5.1.

Fable 5.1 adds beta controls for changing effort per message, applying system instructions to one turn, and showing visible progress updates during long work. These features are useful when one conversation moves between quick routing decisions and expensive analysis, or when users need evidence that a long-running agent is still advancing.

Anthropic also lists content provenance support. Generated text can carry Anthropic's statistical watermark, while supported files created through the Files API can include C2PA credentials. Check the current beta headers and platform availability before depending on these features in production.

The tokenizer is unchanged from Fable 5. Compared with models older than Opus 4.7, however, Anthropic says the same text may produce roughly 30% more tokens. Re-run budget and context tests if you are migrating from an older Claude generation.

The direct Claude API retains Fable 5.1 data for 30 days by default. Zero-data-retention terms apply only to organizations that Anthropic has explicitly approved, so do not assume that an enterprise account automatically has ZDR. Partner-cloud storage and processing terms can differ.

Fable 5.1 is a covered model under Anthropic's commercial terms. Anthropic may route some biology or cybersecurity requests to Opus under its safeguards. When a request is rerouted, customers are not charged Fable 5.1 token prices for the response.

Mythos 5.1 is not an unrestricted public edition. It is limited to Project Glasswing and selected trusted users, with a different safeguard configuration. Standard developers should build against Fable 5.1 and document the possibility of refusals or fallback for sensitive workloads.

A 400 error immediately after changing the model usually points to an incompatible request field, not a bad API key. Check manual thinking configuration, forced tool choice, assistant prefills, and edited conversation history first.

A 401 error usually means the key is missing, malformed, revoked, or loaded from the wrong environment. A 403 can indicate that the account or selected platform does not have model access. A 429 means you reached an account rate limit; apply exponential backoff and check the current usage tier rather than retrying in a tight loop.

Log request IDs and the exact API error body, but redact API keys and sensitive prompt content before sharing logs. For general authentication and rate-limit fixes, see the Claude setup guide for teams.