Claude Fable 5.1: What Changed From Fable 5, Pricing, Cache Reads and API Breaking Changes

What changed in claude fable 5.1?

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 as successors to Fable 5 and Mythos 5. Their API identifiers are claude-fable-5-1 and claude-mythos-5-1.

Both models have a default 1-million-token context window, a maximum output of 128,000 tokens, and always-on adaptive thinking. The September 1 release notes establish the launch specifications, platform availability, and migration changes.

Fable 5.1 vs Fable 5

Anthropic reports improvements in long-session coding, document, spreadsheet and slide work, multistep research, vision, long-context reasoning, and computer use. Its published terminal benchmarks show the following changes:

These are Anthropic-reported results evaluated with production safeguards. They provide evidence for the reported improvements, but they do not establish a specific performance gain for your repository or agent workflow. A migration evaluation should include representative tasks from your own application.

For developers, the release combines capability changes with a narrower pricing change and several compatibility concerns. Treat model quality, token costs, and request compatibility as separate parts of the upgrade.

How Claude Mythos 5.1 differs

The launch announcement describes Fable and Mythos as “the same model, but with different levels of safeguards.”

Fable 5.1 is generally available. Mythos 5.1 has restricted access through trusted programs, with safeguards designed for cybersecurity and life-sciences work; the launch API notes identify access for Project Glasswing participants.

The two models share the listed specifications and token rates. They also differ in conversation-prefix enforcement, which matters when an application edits history containing preserved thinking.

Fable 5.1 pricing: cache reads fall 75%

Fable 5.1 pricing keeps the base input and output rates of Fable 5. The direct price reduction is in cache reads, which fall from $1 to $0.25 per million tokens.

The Fable 5.1 changes documentation lists these rates. All prices below are in US dollars per million tokens.

Cache reads now cost 0.025 times the base input rate. The minimum cacheable prompt remains 512 tokens, and Batch input/output rates are $5 and $25 per million tokens.

What the lower cache rate means

For one million tokens billed specifically as cache reads, the listed charge falls from $1 to $0.25. That is a 75% reduction for that billing category; it is not a 75% reduction in the total request cost.

Input, output, and cache-write rates remain unchanged. When estimating migration savings, keep those categories separate instead of applying the cache-read reduction to the entire bill.

Anthropic estimates 25% lower costs for typical token-billed workloads and up to approximately 45% for highly agentic work. These are workload estimates, not reductions in base input or output prices.

For a broader review of token consumption, see our guide to reducing Claude token usage. For this release, the concrete pricing change to track is how much of your usage is billed as cache reads.

Forced tool choice becomes an API breaking change

Both Fable 5.1 and Mythos 5.1 reject two previously usable tool-choice forms:

• tool_choice: {"type":"any"}
• tool_choice: {"type":"tool","name":"..."}

The result is HTTP 400 with invalid_request_error. The documented error text is:

tool_choice: type "tool" and "any" are not supported for this model.

The migration guide confirms that auto and none remain supported. Validation also applies to Message Batches and token counting, so review those request builders alongside normal message requests.

Updating tool-dependent workflows

The recommended migration uses auto, an explicit instruction to use the relevant tool, and strict: true for schema enforcement. If the purpose of the forced tool was to produce structured JSON, the guide also points to JSON outputs through output_config.format.

These recommendations address different requirements. Schema enforcement concerns the structure of tool arguments; an instruction tells the model which tool you want it to use. Neither should be treated as an unchanged replacement for a request that forces a named tool.

Before switching the model identifier, identify every call site that uses any or a named tool. Review what that call site needs: a tool action, schema-conforming arguments, or a structured response. Then use the migration path appropriate to that requirement.

Preserved thinking needs careful history handling

There are two distinct compatibility issues: moving thinking blocks between model generations, and editing the conversation prefix that precedes a thinking block.

Model compatibility is directional

Fable 5.1 accepts thinking blocks from earlier models. Earlier models cannot read 5.1 thinking blocks.

The API drops incompatible blocks before inference, and dropped blocks are unbilled. The beta header thinking-binding-controls-2026-08-01 exposes these drops through input_transformations.

This matters for model-switching and fallback logic. Moving a conversation to Fable 5.1 and then sending its new thinking blocks back to an earlier model does not preserve those blocks in both directions.

Review model selection separately from conversation storage. Our Claude API model comparison guide provides related context for organizing model choices.

Editing the prefix can invalidate Fable thinking

Changing preceding system, tools, or messages invalidates later Fable 5.1 thinking blocks. An application that rewrites an earlier instruction or modifies the tool definitions must account for the thinking blocks that follow that change.

Default enforcement applies to accounts created on or after August 31, 2026. Older accounts opt in through thinking.block_binding.prefix_mismatch_behavior.

With the binding-controls beta header, "error" rejects invalid blocks, while "drop_block" discards them. The thinking documentation describes these controls.

The recommended history pattern is to preserve assistant turns unchanged and append new messages. Review any code that reconstructs conversations, changes earlier messages, or replaces tool definitions before replaying history.

Mythos 5.1 does not enforce this conversation-prefix check. That difference does not remove its restricted-access requirements.

Launch beta features and their controls

The September 1 release notes include three beta additions relevant to applications that manage long conversations or expose model progress.

Per-message effort is documented at launch on the Claude API. Keep that availability separate from the broader list of platforms that offer the models; model availability alone does not establish support for every beta control.

Turn-scoped system messages use clear_at: "next_user_message". Readable progress updates use thinking.display: "updates". Each feature has its own beta header, so check the configuration and header together when reviewing requests.

For an initial migration, prioritize the required compatibility changes: tool choice and thinking-history handling. Evaluate these optional controls against the needs of your application rather than adding them solely because you changed the model identifier.

Availability and retention at launch

Launch platforms include the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Fable 5.1 is generally available, while Mythos 5.1 access remains restricted through trusted programs.

Anthropic’s undated Fable page lists Pro, Max, Team, and Enterprise availability. Its exact historical publication state is uncertain, so that listing should not be treated as a dated confirmation of each plan’s September 1 availability.

Keep API deployment decisions separate from subscription questions. Our guide to Claude API access versus subscriptions is a useful companion when reviewing how your application will access a model.

Retention requires a separate check

The documented default retention is 30 days. Zero data retention requires express Anthropic authorization.

The launch announcement says eligible customers can use ZDR pending Enterprise Frontier Safeguards, planned for phased availability later in fall 2026. As of September 2, that is a future availability plan.

If retention is a deployment requirement, verify the authorization that applies to your account. Choosing Fable 5.1 or Mythos 5.1 does not, by itself, establish zero data retention.

Key takeaways

• Claude Fable 5.1 launched September 1, 2026, with a default 1-million-token context window, 128,000 maximum output tokens, and always-on adaptive thinking.
• Input and output prices remain $10 and $50 per million tokens. Cache reads fall 75%, from $1 to $0.25.
• Forced tool_choice values any and named tool return HTTP 400; auto and none remain supported.
• Thinking compatibility is directional. Fable 5.1 also checks whether preceding conversation content has changed, with enforcement depending on account creation date and configuration.
• Mythos 5.1 shares the listed specifications and prices, has restricted access, and does not enforce the conversation-prefix check.

FAQ

What changed from Claude Fable 5 to Fable 5.1?

Anthropic reports stronger coding, research, document work, vision, long-context reasoning, and computer use, with higher results on the two published terminal benchmarks. Cache reads are cheaper, and migration requires attention to forced tool choice and preserved thinking.

Are Fable 5.1 input and output prices lower?

No. Input remains $10 and output remains $50 per million tokens, while cache reads fall from $1 to $0.25. Anthropic’s broader savings estimates describe workloads rather than reductions in those base rates.

Why does forced tool choice return a 400 error?

Fable 5.1 and Mythos 5.1 reject tool_choice types any and named tool with HTTP 400 invalid_request_error. The migration guide recommends auto with an explicit tool instruction and strict: true for schema enforcement, or output_config.format for JSON outputs.

Can I switch models without losing preserved thinking?

Fable 5.1 accepts earlier models’ thinking blocks, but earlier models cannot read 5.1 blocks. The API drops incompatible blocks before inference without billing them, and the thinking-binding-controls-2026-08-01 beta header exposes drops through input_transformations.