Citation and evidence

Claude Fable 5.2

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI ModelsAnthropicLarge Language ModelsReasoning Models

Cite this article

Claude Fable 5.2 is a large language model configuration released by Anthropic on November 9, 2026. It succeeds Claude Fable 5.1 and is intended for difficult reasoning, software engineering, research, and long-running agent workflows. Fable 5.2 and Claude Mythos 5.2 share an underlying model, but Fable applies additional safeguards in cybersecurity and biology so it can be offered for general use. Direct Mythos access requires verification.[1]

The model accepts text and images and produces text. Its API supports a 1-million-token context window and up to 128,000 output tokens. Anthropic recommends starting with Claude Opus 5.5 for most workloads, then considering Fable when evaluations show that Opus at higher effort does not meet the task's requirements.[2]

Specifications and access

PropertyDocumented specification
Release dateNovember 9, 2026
Claude API model IDclaude-fable-5-2
InputsText and images
OutputText
Context window1 million tokens
Maximum output128,000 tokens
ThinkingAdaptive, always on
API default efforthigh
Reliable knowledge cutoffJune 2026
Retirement commitmentNot sooner than November 9, 2027

Expanded article table

These are the model's API specifications, rather than a promise that every Claude subscription exposes the full token limits.[2]

The model is available through the Claude API, Claude Platform on AWS, Amazon Bedrock, both Anthropic-managed and Google-managed Claude offerings on Google Cloud, and Microsoft Foundry on Anthropic infrastructure. Bedrock uses the bedrock-runtime endpoint and model identifiers such as global.anthropic.claude-fable-5-2; platform-specific identifiers and feature support should be checked separately.[3]

API pricing

The following USD list prices were checked against Anthropic's documentation on November 17, 2026. All figures are per million tokens; cloud-provider terms, geographic modifiers, and tool charges can differ.[4]

UsageInput or cache priceOutput price
Standard inference$8$40
Message Batches$4$20
Five-minute prompt-cache write$10Not applicable
One-hour prompt-cache write$16Not applicable
Cache read or refresh$0.20Not applicable
Fast mode$16$80

Expanded article table

Batch processing discounts standard input and output by 50%. Cache reads use a 0.025-times multiplier on base input pricing. Cache-write rates apply to tokens written into the cache, rather than replacing the output-token charge. Prices for Fable 5.2 and Mythos 5.2 are the same.[4]

Fast mode

Fast mode changes the inference configuration, rather than selecting a different capability tier. Anthropic advertises up to 2.5 times as many output tokens per second; this refers to generation throughput, not an equivalent reduction in time to the first token or completion time for every task.[5]

On supported Messages API deployments, it is requested with speed: "fast". Microsoft Foundry uses a separate claude-fable-5-2-fast deployment instead. Fast mode has its own rate limits, and a caller can receive HTTP 429 when that capacity is exhausted. It cannot be combined with the Batch API or a Priority Tier commitment. Prompt-cache and data-residency modifiers can apply in addition to the fast-mode rates.[5]

Thinking and effort

Fable 5.2 uses adaptive thinking. In the Messages API, omitting the thinking field or specifying {"type":"adaptive"} enables its supported mode. Disabling thinking, specifying the older enabled mode with budget_tokens, or requesting between_tools returns an error on this model.[6]

Thinking tokens are billed as output and count toward max_tokens, even when their text is hidden. The default display omits readable thinking text; display: "summarized" returns a summary, not the raw internal reasoning. Therefore, hiding thinking is not a way to eliminate its token cost.[6]

The output_config.effort parameter supports low, medium, high, xhigh, and max. The API defaults to high. Anthropic recommends measuring an effort sweep on the intended workload: lower settings can suit simpler work, while xhigh can help difficult agentic tasks. max can consume more time and tokens for uneven gains. Effort is guidance about how much work to do, while max_tokens supplies a hard output limit.[7]

The launch announcement described medium as the default in Claude apps and Claude Code. That product setting should not be confused with the API's documented high default.[1][7]

Reported evaluations

The table gives selected results from Anthropic's November 2026 system card. These are developer-reported evaluations, with different task sets and measurement scales.[8]

EvaluationFable 5.2Fable 5.1Opus 5.5
FrontierCode 2.049.1%32.4%45.5%
SWE-bench Pro93.0%81.2%89.9%
Humanity's Last Exam, with tools74.2%65.6%67.7%
Remote Labor Index41.3%17.9%21.3%
AA-Briefcase v1.1, Elo178516691800
GDPval-AA v2.1, Elo183117631866

Expanded article table

Unless stated otherwise, the system card uses adaptive thinking at max and averages five trials. FrontierCode reports the best effort setting, xhigh for Fable 5.2; Remote Labor Index uses default effort. Context limits vary by evaluation. Fable does not lead every measure: Opus scores higher on the two Elo rows above.[8]

The same card reports that Fable's combined prompt-injection results were weakened by fallback to Opus 4.8, although Fable itself was among the most resistant models tested when it answered the attacks directly. A safety result for the underlying Mythos configuration is therefore not automatically a result for the complete deployed Fable system.[8]

Research and engineering demonstrations

Multi-agent mathematical formalization

Anthropic reported a formalization of Fermat's Last Theorem using 1,000 agents running an early Fable 5.2 version. The engineering account gives a completion time of 28.7 hours. Each agent had a sandbox without internet access, a shell, an editor, Lean and Mathlib; the team shared Git and a PostgreSQL database.[9]

The starting material included 60 theorem statements, definitions, 47 already-proved lemmas from an earlier formalization, and a short sketch of the classical route. This was a reproduction and formal verification exercise building on existing mathematics, rather than a new discovery of the theorem or a proof from no supplied foundations. Anthropic describes it as a toy experiment in scaling agent teams, with command checks and offline monitoring.[9]

Electronic prototypes

In work with Diode, Anthropic tested Fable 5.2 on a speakerphone, a voice-assistant hub, and a handheld virtual pet. It reported that the model prepared circuit layouts, enclosures, parts selections, and firmware, then guided human engineers during assembly and testing. The prototypes worked without changes to their circuitry.[10]

This was a small engineering demonstration, not a commercial product certification. The research report explicitly leaves long-term durability and certification unresolved. It also reports 69% on Diode's internal PCB Gym and 83.5 on BenchCAD 2.1, which measures reconstruction of engineering designs from visual references.[10]

Theoretical physics

The launch associated Fable 5.2 with two theoretical-physics results. The linked preprints describe work on the weak-coupling Hawking-Page transition and the large-central-charge spinless modular bootstrap.[1][11][12]

Their provenance statements identify an internal Claude research model, with physicist checking; the modular-bootstrap account also reports a later independent derivation by Opus 5.5. These publications do not document a general public-model test, and their provenance should be kept separate from claims about the released Fable API.[11][12]

API compatibility and migration

Changing the model ID is only part of migrating an integration. The model-specific migration documentation identifies changes in conversation validation, computer-use tooling, and advisor compatibility.[13]

AreaIntegration consequence
Earlier conversation editsChanging the system prompt, tools, or messages before a bound thinking block can fail the prefix check
Computer useUse computer_toolset_20260801; the older computer_20251124 declaration is rejected outside Amazon Bedrock
Advisor selectionA Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor
Forced tool choicetool_choice values any and tool are unsupported

Expanded article table

Applications should test their actual history-building and model-routing code, rather than assume that an older working integration needs only a new model string.[13]

Preserved thinking

Thinking blocks are tied to the conversation prefix that produced them. A later request must preserve the relevant system instructions, tool definitions, and earlier messages; editing them can produce The block is bound to a different conversation. Passing assistant responses back unchanged and appending new turns preserves the binding.[14]

Blocks also have model and account compatibility limits. Sending a block to a model or account that cannot use it can cause it to be dropped for that request without billing it. Fable 5.2 and Mythos 5.2 can read one another's blocks on the documented supported platforms; other models cannot read Fable 5.2 blocks. Dropping prior reasoning can cause additional generation later, so it has a different cost profile from preserving a usable block.[3][14]

Prompting and agent behavior

Anthropic's prompting guide describes failure modes that remain possible on long jobs: premature stopping, long stretches of tool calls with little visible progress, and excessive delegation in harnesses that encourage subagents. More subagents can raise session costs, so delegation needs evaluation alongside task quality.[15]

The guide recommends explicit completion criteria, instructions about code ownership, and communication expectations when a person is following the work. It also warns that inherited memory or prompt rules can accumulate unnecessary caution, and that instructions that helped older models can have different effects on Fable 5.2. These are reported behavior patterns and integration recommendations, rather than guarantees that a particular prompt will complete every task.[15]

Safeguards and fallback

A classifier refusal is returned as HTTP 200 with stop_reason: "refusal"; it is not necessarily an API transport error. stop_details carries refusal information, although category fields can be null. Applications that inspect only HTTP failures can miss these responses.[16]

The API supports explicit fallback handling, including a beta server-side option on supported platforms and SDK middleware. A fallback is another model answering the request, not Fable's restricted capabilities being enabled. Some refusal categories have no recommended fallback. Batch refusals also require separate handling: server-side fallback is unavailable for batch requests.[16]

Fable's cyber and biology safeguards differ from Mythos's verified-access configuration. An organization requiring the latter should use the applicable Cyber Verification Program or Life Sciences Verification Program, rather than assume that a Fable subscription includes direct Mythos access.[2]

Data retention

Fable 5.2 is a Covered Model requiring 30-day retention. It is not available under zero data retention unless Anthropic expressly authorizes the arrangement. A feature described as ZDR-eligible does not remove the model's own retention requirement.[17]

At launch, Anthropic described Enterprise Frontier Safeguards as a private beta.[1] Its design allows automated safety monitoring with activity data stored in customer-controlled infrastructure. This is an enterprise arrangement, not a default property of Fable API use.[18]

Continue reading

More to read

Topics:AI ModelsAnthropicLarge Language ModelsReasoning Models

References

  1. ^1 ^2 ^3 ^4Anthropic. "Claude Fable 5.2 and Mythos 5.2". November 9, 2026; mathematical-formalization wording amended November 11, 2026.
  2. ^1 ^2 ^3Anthropic. "Claude Fable 5.2: Overview". Claude Platform Docs. Accessed November 17, 2026.
  3. ^1 ^2Anthropic. "What's new in Claude Fable 5.2". Claude Platform Docs. Accessed November 17, 2026.
  4. ^1 ^2Anthropic. "Pricing". Claude Platform Docs. Accessed November 17, 2026.
  5. ^1 ^2Anthropic. "Fast mode". Claude Platform Docs. Accessed November 17, 2026.
  6. ^1 ^2Anthropic. "Thinking". Claude Platform Docs. Accessed November 17, 2026.
  7. ^1 ^2Anthropic. "Effort". Claude Platform Docs. Accessed November 17, 2026.
  8. ^1 ^2 ^3Anthropic. "Claude Fable 5.2 & Claude Mythos 5.2 System Card". November 9, 2026. Executive summary; Sections 1.4-1.6 and 8.1; Table 8.1.A, pp. 207-208.
  9. ^1 ^2Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl, and Lance Martin. "Safely scaling multi-agent teams with Claude Fable 5.2". November 9, 2026.
  10. ^1 ^2Anthropic. "Testing Claude on real-world electronics design". November 9, 2026.
  11. ^1 ^2Anthropic. "The Hawking-Page transition at weak coupling". Hexagon, version 1. November 7, 2026.
  12. ^1 ^2Anthropic. "The spinless modular bootstrap at large central charge". Hexagon, version 1. November 7, 2026.
  13. ^1 ^2Anthropic. "Claude Fable 5.2 and Claude Mythos 5.2 migration guide". Claude Platform Docs. Accessed November 17, 2026.
  14. ^1 ^2Anthropic. "Preserved thinking". Claude Platform Docs. Accessed November 17, 2026.
  15. ^1 ^2Anthropic. "Prompting Claude Fable 5.2". Claude Platform Docs. Accessed November 17, 2026.
  16. ^1 ^2Anthropic. "Refusals and fallback". Claude Platform Docs. Accessed November 17, 2026.
  17. ^Anthropic. "API and data retention". Claude Platform Docs. Accessed November 17, 2026.
  18. ^Anthropic. "Developing Enterprise Frontier Safeguards with our customers". September 1, 2026.

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 2,041 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "Claude Fable 5.2." aiwiki.ai, updated 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/claude_fable_5_2

Edit article