Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, describing them as upgrades to models built for difficult technical work. The company says the pair improves coding, reasoning and vision, and that testers found the models communicate more clearly than earlier versions. Anthropic says Fable and Mythos are the same underlying model, but Fable includes stronger cybersecurity and biology safeguards that allow it to be made available to all users. Organisations pursuing approved cybersecurity or biomedical research can apply for Mythos access through Anthropic’s verification programmes. The distinction is central to the release: capability is shared, but access is governed by the intended use and risk profile.
Fable is broadly available while Mythos remains restricted
Anthropic lists standard pricing at $8 per million input tokens and $40 per million output tokens, down from the previous generation. It says the new models generate outputs up to 30 per cent faster than Fable 5.1 and Mythos 5.1. An optional fast mode can increase token generation speed by up to 2.5 times, at $16 per million input tokens and $80 per million output tokens. The pricing may appeal to developers running long agent workflows, but cost per completed task is more useful than a headline token rate. Context length, output volume, retries, tool calls and human review all affect the final cost of an application.
Anthropic’s published benchmark table reports improvements across several categories. Fable 5.2 scored 49.1 per cent on FrontierCode 2.0 at its highest effort setting, compared with 32.4 per cent for Fable 5.1. It also scored 69.9 per cent on FrontierSWE v2 and 51.4 per cent on AutomationBench. Mythos 5.2 reached 76.3 per cent on Anthropic’s CyScenarioBench cybersecurity evaluation and 89.6 per cent on its CVE Rediscovery test. The company notes that the tasks used by different AI labs are not identical and that results depend on effort settings. These are developer-reported evaluations, not independent guarantees of performance on customer workloads.
Company demonstrations span software, mathematics and scientific research
Anthropic says more than 90 per cent of changes merged into its core codebase are now written by Claude with human guidance. In one experiment, an engineer asked Fable 5.2 to design a prototype operating system from an empty folder. The model coordinated more than 1,000 subagents over eight days to write, test and review the system, which Anthropic says can boot on a laptop and includes drivers, storage, windows, a login screen and an installer. This is a company demonstration rather than evidence that a production operating system can be built without engineers. It does, however, show the model’s ability to coordinate long-running software work.
The research examples are unusually broad. Anthropic says Fable 5.2 solved two open problems in theoretical physics involving black holes and completed a formalisation of Fermat’s Last Theorem in 30 hours using a team of about 1,000 agents. The company also reports that Mythos 5.2 developed a method to predict the structure of whole adeno-associated virus capsids by combining protein-folding models with known symmetry information. In its tests, the method accurately modelled 28 of the 34 capsids it attempted. These results could be valuable if other researchers reproduce them, but they should remain attributed to Anthropic until independent teams verify the methods and outputs.
Safety remains a major part of the release. Anthropic says it ran prerelease evaluations for risks such as sandbox escapes and found the underlying model performed and or better than recent models on most alignment measures. Fable retains the cyber and biology safeguards of its predecessor while producing fewer unnecessary blocks, the company says. Mythos is available for approved cybersecurity work through the Cyber Verification Program, and enterprise customers can access Mythos-powered code scanning through Claude Security. TechCrunch’s coverage of the system card also shows evaluation questions around covert capabilities and whether models recognise when they are being tested. Organisations should assess tool permissions, logging, data retention and human approval requirements before deploying agents for high-impact work. A separate TechCrunch report discusses the system-card concerns around evaluation awareness and model behaviour.