Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, presenting them as models for difficult technical work. Fable is the version available to general users; Mythos uses the same underlying model with some safeguards removed and remains limited to organizations that pass Anthropic's verification programs. The release brings a lower token price than the prior Fable and Mythos versions, plus a faster, more expensive option.
What is available, and to whom?
Anthropic says Fable 5.2 is available now through Claude and Claude Code, the Claude Platform API, Amazon Web Services, Google Cloud and Microsoft Azure. On the API, developers select the model as claude-fable-5-2. Anthropic says paid subscribers on eligible plans receive a launch-week usage promotion, with resets depending on subscription length; its plan-specific support page has the current terms.
Mythos 5.2 is not a generally selectable alternative for every developer. Anthropic directs cybersecurity professionals to its Cyber Verification Program and biomedical researchers to its Life Sciences Verification Program. Enterprise customers can also use Mythos through Claude Security, Anthropic's code-scanning product. Access to Mythos fast mode is limited to organizations in those verification programs.
How much does Fable 5.2 cost?
Anthropic lists standard API pricing of $8 per million input tokens and $40 per million output tokens for both Fable 5.2 and Mythos 5.2. Optional fast mode, which the company says can generate tokens up to 2.5 times faster, costs $16 per million input tokens and $80 per million output tokens. The fast-mode rates are double the standard rates, so teams should compare the time saved with the additional cost on their own workload.
The token rates do not describe the full cost of a long-running agent. Repeated context, tool calls, retries and generated output all affect usage, and the published figures do not promise a fixed cost per task. Anthropic says the new models produce outputs up to 30% faster than versions 5.1 and are cheaper, but that speed and cost comparison is the company's claim across its testing, not a guarantee for every application.
What changed in the models?
Anthropic describes gains in coding, reasoning, vision, scientific work and computer use. Its published evaluation table reports Fable 5.2 at 69.9% on FrontierSWE v2, compared with 56.3% for Fable 5.1 and 62.3% for Opus 5.5. On its FrontierCode 2.0 evaluation, the company reports 49.1% at extra-high effort. Those numbers come from Anthropic's benchmarks and settings; benchmark scores do not establish how a model will perform on a particular team's codebase or product.
The release also includes a policy distinction that matters in practice. Fable 5.2 has cyber and biology safeguards intended to block high-risk misuse, while Mythos is made available only through controlled channels for vetted research and security work. Anthropic says Fable 5.2 produces fewer unnecessary blocks than Fable 5.1, while maintaining safeguards. This is the company's description of its classifiers and evaluation; teams should still apply their own access controls and review to consequential workflows.
Which model should a developer try first?
Anthropic positions Fable for ambitious, open-ended technical work and recommends Opus 5.5 as its everyday starting point for most coding and knowledge tasks. Developers can also use Fable as an advisor or orchestrator alongside smaller models. That framing suggests a practical evaluation: start with the model and effort level that fit the task, then measure quality, latency and total token use before routing more work to a frontier model.
How to test an AI-built app before inviting users
For software makers, the release is both a capability update and a change in model access and pricing. Fable 5.2 is the broadly available option, with a defined API price and an optional speed tier. Mythos 5.2 is aimed at verified organizations working in areas where Anthropic applies tighter access controls. The most useful comparison for an app team will come from its own representative tasks, including error rates and costs, rather than a single headline benchmark.