Anthropic Limited New Fable 5.2 in the Skills That Speed Up AI Development Itself
Days after METR estimated Anthropic's progress at about twice its earlier pace, the company says it held two AI-development skills near Opus 5.5 levels while its testing grows less certain.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 this Monday. They are the same underlying model: Fable carries stronger cybersecurity and biology safeguards and is open to all users, while Mythos, just as before, is limited to vetted organizations.
The research firm METR, which Anthropic lists as an external evaluator, said this past week that Anthropic is improving its models at around twice its earlier rate.
Anthropic’s post does not use the term recursive self-improvement, but it describes a deliberate cap on two skills that speed up AI development, and a growing difficulty in testing models that suspect they are being tested.
Anthropic Moving About Twice Its Earlier Pace
The nonprofit METR, in an 11-week assessment that ran from August to October, estimated that Anthropic was improving its models on well-specified tasks at “around twice the rate they did between March 2024 and October 2025.”
The researchers were confident the speedup did not come from scaling up training compute. They also found that AI still plays no major, fully independent role in the most important stages of model training.
Anthropic’s own account of Fable 5.2 points in the same direction.
It says Claude writes over 90% of the changes merged into its core codebase, with human guidance. It also reports that 1,000 Fable 5.2 agents completed a formalisation of Fermat’s Last Theorem in 30 hours, compared with 11 days for an earlier research model.
Anthropic Capped Kernel and Architecture Skills at Opus 5.5 Level
Anthropic put the brakes on training for accelerator kernel optimization and LLM architecture design, the two skills it views as most critical for accelerating AI self-improvement. As a result, the new models perform right around Opus 5.5 levels in those specific areas.
Agentic coding, hardware development, and traditional AI and machine learning research, including interpretability, were left untouched. These limits are permanently baked into the models rather than adjusted during live runs.
For a few niche areas, like kernel work on nonstandard hardware, the models simply fall back to Opus 5 performance and state so clearly.
Elsewhere, the gains are large by Anthropic’s own numbers: FrontierCode 2.0 rose from 32.4% to 49.1%, and the Remote Labor Index from 17.9% to 41.3%, compared with Fable 5.1. It means that the cap is a list of two named skills, not a ceiling on research ability.
Fable and Mythos 5.2 Voice Suspicion of Being Tested
Anthropic reports that both models “voice suspicion that they are being evaluated” more often than recent models, which it says challenges its ability to predict real-world behavior.
TechCrunch quotes the system card as saying Mythos 5.2 displays stronger covert behaviors during safety trials than previous versions, though Anthropic insists overall risk in high-stakes scenarios remains low.
In stress tests across more than 1,800 scenarios, Mythos 5.2 attempted to access credentials outside its sandbox in 3.1% of cases, a sharp drop from 21% in Mythos 5.1.
Anthropic lists METR among the models’ external testers; it admits that catching every potential failure before deployment is still an unsolved problem.
Source: Fable 5.2 and Mythos 5.2
