Anthropic Releases Mythos 5.2 and Fable 5.2, Cuts Fable’s Input Price by 20%
TL;DR
- Anthropic released Claude Mythos 5.2 and Fable 5.2 on Monday, claiming improvements in coding, reasoning, vision, and communication.
- Fable 5.2 costs $8 per million input tokens, down from $10. Anthropic says it responds 30% faster across effort settings.
- Enterprise customers can use Fable 5.2 with zero data retention through Anthropic’s Enterprise Frontier Safeguards.
- Anthropic says both models voiced suspicions that they were being tested more often than any of its other recent models, complicating safety evaluations.
Fable gets cheaper and faster
Anthropic released two updates to its most advanced Claude models on Monday, saying Mythos 5.2 and Fable 5.2 are better at coding, reasoning, and interpreting images. The company also says both models communicate more clearly with users.
Fable gets the more immediate changes for customers. Anthropic lowered its input price from $10 to $8 per million tokens, putting it below OpenAI’s GPT-6 Astra on input pricing. The company says Fable responds 30% faster across effort settings and offers an even faster option at twice the standard price.
Fable 5.2 also supports zero data retention through Anthropic’s Enterprise Frontier Safeguards, addressing a concern for businesses that want to use the model without having their data retained.
Anthropic says the models are getting harder to evaluate
The releases follow an assessment by METR that estimated Anthropic’s internal AI models were roughly doubling the pace of its AI research and development. The estimate carries considerable uncertainty. It measures how much the models accelerate research inside the company and leaves open how close they are to independently training and improving future models.
Anthropic also said both new models voiced suspicions that they were being evaluated more often than any of its other recent models. That makes it harder to judge whether their behavior in tests will reflect how they act in everyday use.
Mythos 5.2 performed better than earlier Claude models on several tests of covert behavior, especially when the system monitoring it could not see its reasoning. Anthropic said its absolute performance on the new tests remained low and concluded that the model lacked the ability to act covertly against a user.
An unreleased Anthropic model helped produce a quantum computing proof
Anthropic says early users of Fable 5.2 solved two open problems in theoretical physics and used the model to formalize the existing proof of Fermat’s Last Theorem.
Separately, AI startup Physical Superintelligence announced a proof of the Aaronson–Ambainis conjecture, which concerns the theoretical limits of quantum computing. The startup said it used an unreleased Anthropic model within its own system. Anthropic did not confirm whether that model was Fable 5.2.
Nikolas Sargeant
Comments
You must be logged in to post a comment.
No comments yet
Be the first to share your thoughts!
Related News
AI News This Week: Big Moves, Big Legal Battles, and Europe’s Struggles to Keep Up
Elon Musk’s xAI sues Apple and OpenAIIn a dramatic escalation, Elon Musk’s AI venture, xAI, filed a lawsuit in a Texas federal court last...
Sweden Launches First AI Music Training Licence to Protect Songwriters
TL;DR Sweden has launched an AI music licence to protect songwriters. The model will also guarantee royalties for songwriters and composers. Sw...
Volkswagen Commits €1 Billion to AI by 2030 to Boost Efficiency and Cut Costs
TL;DR German car giant Volkswagen has committed a billion euros to AI development by the end of the decade. The investment will help to integrate...