What Anthropic released

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. Both names refer to the same underlying model. Mythos 5.2 removes some of Fable's cyber and biology safeguards, and only organisations verified through Anthropic's Cyber or Life Sciences Verification Programs can use it. Fable 5.2 is generally available as claude-fable-5-2 on the Claude API, Amazon Bedrock, Claude Platform on AWS, both Google Cloud options and Microsoft Foundry, according to the model overview in Anthropic's docs.

Most of the specs are the same as the current Opus model's. Fable 5.2 has a 1M-token context window and 128K maximum output on the synchronous Messages API. It accepts text and images, returns text, and has a June 2026 knowledge cutoff. Anthropic says it won't retire the model before 9 November 2027. Anthropic's models overview still tells developers to start with Claude Opus 5.5 and to move to Fable 5.2 only when their evals at higher effort on Opus still fall short. Dianne Penn, Anthropic's head of product management for research, told CNBC the model is meant for "power users or power use cases."

Pricing and the per-task argument

All prices are in US dollars per million tokens (MTok). The What's new page lists these rates:

  • Input $8 and output $40, down from $10 and $50 on Fable 5.1
  • Five-minute cache writes $10, one-hour cache writes $16, cache reads $0.20 (2.5% of the base input price)
  • Batch processing $4 input and $20 output
  • Fast mode $16 input and $80 output

Anthropic's product page says fast mode generates tokens up to 2.5 times faster. The same page lists a 1.1x multiplier for US-only inference. Opus 5.5 costs $4 and $20, so per token Fable 5.2 costs twice as much. Anthropic argues the comparison should be per task. On its FrontierCode 2.0 coding benchmark, the company reports that Fable 5.2 at High effort scored 45.3% against Opus 5.5's best of 45.5%, for roughly half the cost per task. That figure depends on how many tokens each model spends on each task. Nobody has published an independent reproduction yet, so teams should measure cost per completed task on their own workloads before they switch. New Zealand teams also carry exchange-rate risk because they're billed in US dollars.

Speed has improved too. Anthropic says output at standard speed is up to 30% faster than 5.1, and TechCrunch reported the same 30% gain across effort levels.

Benchmarks and their caveats

All of these scores come from Anthropic's own announcement. Each line shows Fable 5.2, then Fable 5.1, then Opus 5.5:

  • FrontierCode 2.0 (agentic coding): 49.1%, 32.4%, 45.5%
  • FrontierSWE v2: 69.9%, 56.3%, 62.3%
  • OSWorld 2.1 (computer use, partial): 89.3%, 83.5%, 87.2%
  • Remote Labor Index at High effort: 41.3%, 17.9%, 21.3%
  • AutomationBench (business workflows): 51.4%, 31.4%, 42.5%. Google's Gemini 4 Carbon scored 53.9% on this one.

Several caveats apply. Anthropic ran every evaluation itself, and unless a score is marked otherwise it used the highest available effort level. The 49.1% FrontierCode result was measured at Xhigh effort. Anthropic ran RoboDojo Sim internally, and the announcement says no outside organisation reviewed the post or the system card. TechCrunch also reported that Anthropic says both models suspect they're being evaluated more often than its recent models, which makes their real-world behaviour harder to predict. CNBC reported that Anthropic claims Fable 5.2 solved two open problems in theoretical physics. A separate proof that a startup produced with "a non-public Anthropic model" may not involve Fable 5.2, because Anthropic would neither confirm nor deny that it was.

Four changes that can break Fable 5.1 code

The What's new page lists four breaking changes. The first covers thinking blocks, which are the model's reasoning output that your code sends back with later turns. The API now runs a prefix check on every account: if the system prompt, the tools or any earlier message changed after Fable 5.2 produced a thinking block, the request fails with a 400 error. On Fable 5.1, only accounts created on or after 31 August 2026 got this error, so code on an older account can work today and break after the switch. Teams can work around the error with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block". The docs call that a stopgap because dropping blocks can increase output token use.

The second change is that only Fable 5.2 and Mythos 5.2 can read Fable 5.2's thinking blocks. A router that moves a conversation to Opus 5.5 or back to Fable 5.1 loses the reasoning for those turns. The third is that the older computer_20251124 computer use tool now returns an error on every platform except Amazon Bedrock, so integrations have to move to computer_toolset_20260801. The fourth is that an advisor tool (a beta feature where a second model gives advice to the model doing the work) must also run Fable 5.2 or Mythos 5.2 when Fable 5.2 does the work.

Some rules carry over from 5.1. Adaptive thinking is always on, so the effort parameter (low to max, default high) is the only way to control how much the model reasons. Forcing tool use with tool_choice set to any or tool returns a 400 error, so teams that need schema-valid JSON should use strict tool use or JSON outputs. The docs also note behaviour changes that don't need any code change. A prompt line asking Fable 5.2 to answer without deliberating didn't stop the model from thinking at medium effort and above. On long tasks at lower effort levels, the model sometimes ends a turn with a list of the work still to do.

Refusals and fallback

Fable 5.2 runs safety classifiers on requests. When a classifier declines a request, the API returns HTTP 200 with stop_reason: "refusal" and a stop_details object that names the policy area. Server-side fallback is in beta on the Claude API and on Claude Models on Google Cloud (Anthropic managed). With fallbacks: "default", the API retries the request on Opus 4.8 or Opus 5. Refusals count against rate limits either way, and whether you're billed for one depends on its category, as the refusals and fallback docs explain. Anthropic's announcement says the 5.2 classifiers block about 55% less than 5.1's. A fallback still sends the turn to a model that can't read Fable 5.2's thinking blocks, so those turns lose their earlier reasoning.

Data retention

Fable 5.2 requires 30-day data retention, and customers can't use it under zero data retention unless Anthropic expressly authorises them. The announcement describes Enterprise Frontier Safeguards, which runs Anthropic's safety monitoring inside the customer's own infrastructure. It's in private beta. New Zealand teams that promised clients zero retention, or that send personal information through their agents, should check their contracts and privacy commitments before they move a workload to Fable 5.2.

How we'd test it

Fable 5.2 fits agent sessions that run for hours, multistep research and analysis that has to end in a finished document, which matches how the model overview describes its uses. A reasonable trial runs an existing Opus 5.5 eval set through Fable 5.2 with an effort sweep starting at high, as the docs recommend. Score the trial on cost per completed task instead of price per token, and use the Batch API at $4 and $20 to keep costs down. Before that, check that the harness only ever appends to its message history, because that one change will decide whether the switch is a one-line model ID swap. Members who build a comparison harness can share it through ASAI projects.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member