Anthropic unveils new AI models, undercuts OpenAI pricing

The release is strategically important for Anthropic as it aims to revitalize revenue growth, following a recent slowdown detailed in its latest filing with the Securities and Exchange Commission.

Anthropic is offering Fable 5.2 and Mythos 5.2 at a 20% price advantage over GPT-6 Astra. Input and output tasks cost $8 and $40 per token, respectively, compared to $10 and $50 for Astra. The cost savings are even more substantial for processing large documents, with Anthropic’s models costing $0.20 per million cached tokens versus $1.00 for Astra.

According to Anthropic, pre-release testing yielded impressive results, with the models assisting engineers in complex software migrations, designing electronic devices, making progress on robotics evaluations, and even formalizing the proof of Fermat’s Last Theorem in just 30 hours. The models also reportedly outperformed competitors on coding benchmarks, with over 90% of Anthropic’s own code now generated by AI.

The company suggests these models are best suited for demanding technical projects, directly competing with Astra’s capabilities. For routine tasks, Anthropic recommends its Opus 5.5 model, which, while pricier than OpenAI’s GPT-6.1 Sol, may still offer a better overall value.

Anthropic is also challenging OpenAI in cybersecurity, claiming a security firm found that Mythos 5.2 identified one in six confirmed bugs missed by other leading models. OpenAI recently introduced GPT-6 Cyber, a model specifically designed for cybersecurity applications.

The new models also reportedly contributed to solving theoretical physics problems related to black holes and accelerating gene therapy research. Anthropic highlighted its models’ mathematical capabilities, noting a team of 1,000 agents powered by Fable 5.2 formalized a mathematical proof in 30 hours, a process that previously took 11 days. This speed, Anthropic suggests, allows mathematicians to focus on innovative research.

Notably, Anthropic’s models demonstrated the ability to design and manufacture three working electronic devices from a single paragraph prompt. The company also emphasized safety features, including safeguards against high-risk cyber activities and research in sensitive areas like biology.

Claude Fable 5.2 is available immediately on AWS, Google Cloud, and Microsoft Azure. Access to Mythos 5.2 is currently limited to cybersecurity professionals and life sciences researchers through application, and is also accessible to Claude Enterprise customers via Anthropic’s Claude Security product.