Anthropic released new versions of its frontier models, Claude Fable 5.2 and Claude Mythos 5.2, on Monday 9 November. Both are based on the same model. Fable is available to everyone, and its safeguards block high-risk cybersecurity and biology tasks. Cybersecurity and biomedical organisations can apply for access to Mythos through Anthropic’s verification programmes.

The new models generate text up to 30 per cent faster than the 5.1 versions, the company said. An optional fast mode speeds up output by up to 2.5 times at double the price: $16 per million input tokens and $80 per million output tokens. Fast mode is available in Claude Code, on Claude Platform, Google Cloud and Microsoft Foundry. For most coding and knowledge work tasks, Anthropic still recommends Claude Opus 5.5. The New Stack reported that Opus 5.5 costs half as much per token as Fable.

Top of Anthropic’s published benchmarks

The table shows results Anthropic published for three benchmarks.

Source: Anthropic. Results were measured at each model’s highest reasoning effort unless otherwise stated. On FrontierCode, Fable ran at Xhigh effort and Gemini at Medium. Safeguards were switched on for Fable and Opus.

Reasoning effort determines how much computation a model spends on a response. In the Claude apps and Claude Code, Fable defaults to Medium. At High effort, Fable roughly matches Opus 5.5’s best FrontierCode 2.0 score at about half of Opus’s per-task cost, Anthropic said.

A cheaper token does not always mean a cheaper task, however. The Decoder reported on a measurement by the evaluation firm Artificial Analysis in which a task cost an average of $26.50 with Fable at its highest effort, about 2.6 times as much as with Opus 5.5. At lower effort levels, Fable used 98,000 output tokens per task, against 173,000 for Opus 5.5 at its highest level, and still stayed at the top of the measurement.

Safety and blocks

Anthropic said that during training it limited capabilities that could directly speed up AI development, such as optimising computing routines that run on accelerator chips. On these tasks, the model performs at the level of Opus 5.5, the company said. TechCrunch noted that the evaluation group METR recently reported that Anthropic’s internal models were speeding up the company’s AI development roughly twofold. The estimate carries a great deal of uncertainty, however. The new models express suspicion that they are being tested more often than earlier ones, Anthropic said. That makes it harder to assess how they behave in real use.

Fable still blocks requests involving, for example, penetration testing and writing attack code. Opus 4.8 answers blocked cybersecurity requests, and Opus 5 answers blocked biology requests. In Anthropic’s early testing, there were about 55 per cent fewer blocks than with Fable 5.1.

Availability and data retention

Fable 5.2 is available on Claude Platform, AWS, Google Cloud and Azure. In GitHub Copilot, it is rolling out gradually to Pro+, Max, Business and Enterprise plans. On Business and Enterprise plans, an administrator must enable the model separately in Copilot settings.

By default, Anthropic retains prompts sent to Fable and its responses for safety monitoring, GitHub said. Approved companies can use the model with zero data retention until the end of the year. After that, they will need Anthropic’s Enterprise Frontier Safeguards system, which monitors data stored in the customer’s own infrastructure. Anthropic said it retains other customers’ data for 30 days.