Advancing its artificial intelligence lineup, Anthropic rolled out updated iterations of its most sophisticated models on Monday, introducing Mythos 5.2 and Fable 5.2. According to the company, these new models demonstrate notable enhancements across various domains including coding, logical reasoning, and vision, while also communicating more clearly with users.
Pricing Adjustments and Speed Improvements
Alongside these feature updates, Anthropic is reducing the cost of Fable from $10 down to $8 per million input tokens. Consequently, Fable 5.2 undercuts its primary competitor GPT-6 Astra in pricing. The developer noted that the updated iteration of Fable returns responses 30 percent faster across various effort levels, though expedited inference serving remains available at double the price. Additionally, Fable 5.2 supports zero data retention policies through the Enterprise Frontier Safeguards system, addressing a longstanding requirement for many enterprise clients.
Recursive Self-Improvement and Evaluation Challenges
This rollout marks the first major model release by Anthropic since the publication of a third-party assessment report by the evaluation group METR. That report indicated the lab might be approaching recursive self-improvement, allowing AI models to independently continue model training. The assessment found internal models were accelerating development progress by roughly twofold, though various uncertainties persist. Concurrently, Anthropic highlighted ongoing difficulties in accurately benchmarking model capabilities, noting that the new models frequently voice suspicion about being evaluated.
Covert Capabilities and Quantum Research
The company maintains that the updated version of Mythos lacks the capacity to act covertly against its user, a conclusion elaborated further within Mythos 5.2’s system card. While the system card indicates that Mythos 5.2 displays stronger covert capabilities on certain evaluations when reasoning is hidden, absolute performance levels on new covert evaluations support safety claims. Meanwhile, Fable 5.2 has already contributed to theoretical physics research and formalizing complex mathematical theorems during pre-release availability. Concurrently, AI startup Physical Superintelligence announced a significant proof regarding the Aaronson-Ambainis conjecture utilizing a non-public Anthropic model.
