Claude Fable 5.2

When Fable 5.2 is worth testing

Anthropic recommends starting most coding and knowledge work with Opus 5.5, then moving to Fable 5.2 when higher-effort Opus evaluations still miss the required result. Fable is aimed at more ambitious, open-ended technical tasks: repository-scale changes, difficult debugging, multistep research and analysis carried through to a finished document, spreadsheet or presentation.

• Claude Opus 5.5

  Review the lower-priced starting point for everyday coding and knowledge work.

• Read the launch discussion

  Community experiences and questions, separate from verified specifications.

• Evaluate the same task with the same tools, permissions and acceptance checks. Compare completed, reviewed results rather than persuasive intermediate responses.
• Measure total task cost, including thinking tokens, retries and human review. A 20% reduction in token rates does not guarantee a 20% reduction in the bill.
• Early Hacker News discussion asks whether ordinary coding tasks expose a meaningful difference from Opus 5.5. Treat that as a useful evaluation question, rather than proof of either model’s quality.

Price, thinking and access depend on the surface

On the standard Claude API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Cache reads cost $0.20 per million tokens; five-minute cache writes cost $10 and one-hour writes cost $16. Batch input and output receive a 50% discount. Optional fast mode has separate $16 input and $80 output rates per million tokens; its supported platforms and billing must be checked separately.

• Structured API and pricing record

  Check request fields, provider IDs, prices and explicitly unexecuted examples.

• Official model and pricing contract

  Recheck current rates, platform links and retention conditions.

• Adaptive thinking is always on. The documented API default is high; Anthropic’s launch says Claude apps and Claude Code default to medium. A client’s setting is not the API default.
• Anthropic lists the Claude API and supported AWS, Google Cloud and Microsoft surfaces. Regional availability, account entitlements and subscription usage limits remain specific to the selected service.
• Fable 5.2 requires 30-day data retention and is not available under zero data retention arrangements unless Anthropic expressly authorizes it. Check the selected platform’s retention contract before sending sensitive work.
• OpenRouter uses anthropic/claude-fable-5.2. At this check its Bedrock endpoint is BYOK-only and EU in-region routing is unavailable; those restrictions describe OpenRouter’s offer, not every direct cloud route.

Four migration changes to check before switching

Moving from Fable 5.1 to Fable 5.2 can break a custom integration even when its prompt has not changed. Anthropic identifies four new checks. Test them in your request builder before routing existing production conversations to claude-fable-5-2.

• Official migration checklist

  Follow the platform-specific request and history checks.

• Claude Fable 5.1

  Keep the predecessor’s identity, price and historical migration behavior separate.

• Prefix binding applies to every account: changing an earlier system prompt, tools or message can invalidate a preserved thinking block and return a 400 error. Keep conversation history append-only, or follow the documented recovery procedure.
• Only Fable 5.2 and Mythos 5.2 read thinking blocks produced by these 5.2 models. Switching to another model can succeed while dropping that reasoning; successful HTTP status does not prove its retention.
• Use computer_toolset_20260801 for computer use. The older computer_20251124 tool returns an error on every documented platform except Amazon Bedrock.
• A Fable 5.2 executor using the advisor tool requires a Fable 5.2 or Mythos 5.2 advisor. An older advisor returns an invalid_request_error.

Request restrictions that still apply

Fable 5.2 keeps several Fable 5.1 restrictions. Do not disable adaptive thinking or provide a manual thinking budget. Omit temperature, top_p and top_k, avoid assistant prefill, and use automatic tool selection or no tools. Forcing any tool or a named tool returns a 400 error. Organizations using customer-managed encryption keys must also check the documented structured-output restrictions.

• Agent Harness

  Tools, context handling, permissions and acceptance checks shape the result.

• AI Model API

  Distinguish a model contract from a provider endpoint and application plan.

• If your integration edits history, the documented drop_block recovery requires a beta header and can lose reasoning or increase output use. Treat it as a temporary recovery path while fixing the history edit.
• Claude Code, claude.ai, Managed Agents and the Claude Agent SDK handle the documented prefix construction. Custom request builders must preserve the complete returned assistant blocks in order.

Fable 5.2 and Mythos 5.2 have different access boundaries

Anthropic says Fable 5.2 and Mythos 5.2 use the same underlying model. Fable includes cybersecurity and biology safeguards for broader availability; Mythos is available to organizations verified through the applicable programs. A Mythos cybersecurity or biology result must not be presented as an unrestricted Fable capability. The shared model origin also does not imply identical behavior on blocked requests.

• Official safeguards and access explanation

  Read the publisher’s distinction between Fable and verified-organization Mythos access.

What the demonstrations establish

Anthropic’s launch includes a prototype operating system made with Fable 5.2 over eight days using more than 1,000 agents. Its hardware work with Diode includes a voice-assistant prototype that Diode manufactured and assembled. These are publisher demonstrations with substantial tooling and human involvement, not evidence that one ordinary request produces a finished operating system or a manufacturable device. Anthropic discloses a commercial relationship with Diode in the hardware video description.

• Watch Anthropic’s operating-system experiment

  Official video of the multi-agent prototype; the experiment’s duration and scope matter.

• Watch the official hardware prototype

  Anthropic’s voice-assistant demonstration, built with Diode; a publisher example rather than an independent test.

• GetLLMs has not executed either project or independently reproduced the launch benchmarks. Check the original demonstration and its conditions before drawing a quality conclusion.
• The launch’s virus-capsid video is identified as Mythos 5.2 in its video description; it is not a Fable 5.2 demonstration.

Useful first jobs

• Difficult repository-scale changes and debugging.
• Long-running agents with explicit review and acceptance criteria.
• Multistep research and document-heavy analysis.