Anthropic describes Claude Fable 5.2 and Claude Mythos 5.2 as the same underlying model with different safeguards and access. Fable is offered through general Claude channels with cybersecurity and biology safeguards; Mythos is limited to vetted organizations pursuing eligible cybersecurity or biomedical work. Notebookcheck reported that Fable 5.2 was released on November 9, 2026.
How Fable and Mythos differ
The practical distinction is who can access each configuration and what safeguards apply. Anthropic says Fable can block high-risk cybersecurity work, including penetration testing and exploit generation, as well as most research-biology tasks. For blocked requests, Anthropic says the system may fall back to Opus 4.8 for cybersecurity or Opus 5 for biology. Mythos has fewer of those safeguards, but access is restricted through Anthropic’s verification programs.
| Aspect | Claude Fable 5.2 | Claude Mythos 5.2 |
| Underlying model | Same model Anthropic describes for Mythos 5.2 | Same model Anthropic describes for Fable 5.2 |
| Safeguards | Includes safeguards for high-risk cybersecurity and most research-biology tasks | Does not include certain Fable safeguards |
| Access | Offered through general Claude channels | Limited to vetted organizations through cybersecurity or life-sciences verification programs |
| Intended work | Ambitious, open-ended technical tasks; Anthropic recommends Opus 5.5 for most coding and knowledge work | Eligible specialist cybersecurity or biomedical research |
Anthropic lists Fable on Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Its announcement also lists fast mode for Claude Code and Microsoft Foundry. That does not change Mythos’s verification requirement.
NeoTeo’s earlier coverage of Fable 5.1 and Mythos 5.1 covered the previous model pair; Anthropic describes the 5.2 versions as sharing a model but retaining different access and safeguards. Anthropic’s announcement sets out those terms.
What Anthropic’s performance claims show
Anthropic reports that Fable 5.2 and Mythos 5.2 generate outputs up to 30% faster than their respective 5.1 models. That is the company’s stated maximum, not a promise that every task will finish 30% sooner.
The company’s benchmark results are not a sweep across every test. In Anthropic’s FrontierCode 2.0 results, Fable 5.2 scored 49.1% at Xhigh effort; Fable 5.1 scored 32.4% at the same effort. Anthropic lists Opus 5.5 at 45.5%, GPT-6 Astra at 42.8%, and Gemini 4 Carbon at 41.1% at Medium effort, so those cross-model comparisons use different effort settings. On AutomationBench, Gemini 4 Carbon scored 53.9%, ahead of Fable 5.2’s 51.4%. Anthropic says its table reflects its own evaluations and that effort settings vary where noted.
Anthropic also reports a technical demonstration in which 1,000 Fable agents formalized Fermat’s Last Theorem in 30 hours with minimal human intervention. The company compared that with an earlier 11-day effort using an unreleased research model, a purpose-built harness and human guidance. Anthropic says the Fable effort used one-third as many action tokens and produced a proof seven times shorter; the company amended its account on November 11 to clarify what information the model received.
For Mythos, Anthropic reports that its method accurately modeled 28 of 34 attempted natural and engineered capsids. The result describes the company’s modeling tests, not clinical outcomes.
A separate walkthrough by The AI Advantage shows browser-based interactive games and presentation materials attributed to Fable 5.2. Those examples give a sense of the kinds of outputs shown in the video; they are not a general performance measure.
Anthropic recommends Opus 5.5 as the starting point for most coding and knowledge work, reserving Fable or Mythos for more ambitious, open-ended technical tasks. That guidance fits the distinction between a broad-use model with safeguards and a restricted configuration for eligible specialist work.
The benchmark lead is task-dependent: check the evaluation and effort setting before treating a score as a guide to your own workload.
API rates, the launch offer and user reports
Anthropic lists the same standard API rates for both models: $8 per million input tokens and $40 per million output tokens. Optional fast mode is listed at $16 per million input tokens and $80 per million output tokens, with token generation up to 2.5 times faster. The speed is a stated maximum; the higher rates apply to fast mode.
Anthropic says its launch promotion includes Fable for up to half of the weekly usage limits on eligible Pro plans and standard Team seats. Eligible plans or seats receive one to three resets depending on subscription tenure. Notebookcheck reported that the temporary Pro offer was scheduled to end November 16, 2026, at 11:59 p.m. California time. Notebookcheck also reported that an eligible reset restored Fable usage already spent and refilled the five-hour window.
Individual Reddit users describe different results. One reported using a Pro five-hour window in less than one question; another said Fable made little difference on their science prompts, while a different user reported that it solved difficult bugs that Opus 5.5 had not, but found it slow and costly. These are accounts of particular workloads, not a typical quota rate or a controlled comparison. (Quota and task-use reports; contrasting task experiences.)
Anthropic makes Claude Fable 5.1 generally available, limits Claude Mythos 5.1 to vetted users
