Anthropic debuts updated Claude Fable and Mythos models, priced 20% below OpenAI’s GPT-6 Astra

The new models, which could be Anthropic’s last big releases before its IPO, claim to offer higher performance than Astra at a lower price point—a key factor for earning and retaining customers in a world where professionals seem quick to switch to models based on cost, as long as the performance holds up.

The launch is critical for Anthropic as it seeks to re-accelerate revenue growth, following a third quarter slowdown revealed in latest S-1 filing with the SEC.

Fable 5.2 and Mythos 5.2 are 20% cheaper than Astra for ordinary input and output tasks, at $8/input token and $40/output token, compared to $10 and $50, respectively, for GPT-6 Astra. For prompts with large amounts of information, such as documents, the cost advantage is more pronounced. Anthropic’s new models are 80% cheaper, at $0.20/million cached tokens compared to $1.00/million cached tokens for GPT-6 Astra.

Anthropic reports impressive results from its pre-release testing. The new models “helped our engineers complete otherwise-infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in the physics of black holes, made substantial progress on robotics evaluations, and formalized the proof of Fermat’s Last Theorem in 30 hours—among many other applications,” the company said.

Fable 5.2 and Mythos 5.2 perform better than its competitors on coding evaluation benchmarks, Anthropic said, noting that over 90% of the company’s own code is written by AI. While Anthropic gained an early lead among programmers, OpenAI’s Codex has gained ground in recent months due to its lower pricing, The Information reported in late October. The combination of lower pricing plus superior performance, if the company’s claims hold up in real-world testing, could help win them back.

The new models are best when used for the “most ambitious, open-ended technical work,” Anthropic said. In other words, they’re competing with Astra at the frontier. However, for daily tasks, Anthropic recommends its customers use its more cost effective model, Opus 5.5. That model is more expensive than its direct competitor from OpenAI, GPT-6.1 Sol, although it still could end up being a better deal if it yields a high quality result more often. That’s for users to decide as they inevitably begin to try Anthropic’s latest releases and weigh options on a daily basis.

Anthropic is also trying to compete with OpenAI on cybersecurity tasks, such as finding and fixing vulnerabilities in code. The company claims an unnamed security company “found that one in six confirmed bugs found by Claude Mythos 5.2 were missed by all of the other frontier models they tested.” Late last week, OpenAI debuted a cybersecurity-focused model, GPT-6 Cyber.

The new models also reportedly solved two theoretical physics problems related to black holes, and identified a way to speed up research into gene therapies. In light of OpenAI making headlines for solving mathematics proofs that have eluded academics for decades—most notably the 90-year-old Navier–Stokes Millennium Prize Problem—Anthropic is highlighting its new models’ mathematics chops as well.

A team of 1,000 agents powerd by Fable 5.2 produced a formalization, or translation of a mathematical proof into a format a computer can check step by step, in 30 hours, compared with 11 days for a previous, unreleased research version of Claude. In a subtle jab at the mathematics community’s backlash to OpenAI pouring millions of dollars into solving proofs and upending the field, Anthropic notes that its quick formalizations can help human mathematicians check their proofs faster so they can “spend more time exploring novel ideas.”

One unique area Anthropic seems to be leaning into more than OpenAI is hardware and robotics. The new models were allegedly able to design and produce three working electronic devices, complete with circuit boards, 3D-printed enclosures, and software, based on a one-paragraph prompt.

Finally, Anthropic highlighted the models’ safety features, including safeguards that block “high-risk cyber activities including penetration testing, exploit generation, and binary-based vulnerability, as well as most tasks in research biology.”

Claude Fable 5.2 is available today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Mythos 5.2 access is application-only for cybersecurity professionals and life sciences researchers. Claude Enterprise customers can also access Mythos 5.2 through Anthropic’s Claude Security product.