Anthropic releases Claude Fable 5.2 and Mythos 5.2

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Nov. 9, the company announced, pitching the pair as its most capable models for coding, finance and the sciences.

The two share one underlying model. Fable 5.2 carries cybersecurity and biology safeguards and is open to all users. Mythos 5.2 strips those safeguards and is restricted to vetted organizations through Anthropic’s Cyber and Life Sciences Verification Programs. Both cost $8 per million input tokens and $40 per million output tokens, down from $10, and generate output up to 30% faster than Fable 5.1, according to Anthropic. An optional fast mode speeds generation by up to 2.5 times at double the price.

The release lands four days after METR, a nonprofit that evaluates frontier AI models, published a report finding that Anthropic’s internal models are now accelerating its own AI research by roughly a factor of two, TechCrunch reported. During prerelease testing, Fable 5.2 also formalized the proof of Fermat’s Last Theorem in 30 hours and solved two open problems in the physics of black holes. Topping a coding benchmark is a quarterly event. Producing preprints that working physicists read is not, and Fable 5.2 has done both.

On FrontierCode 2.0, an agentic coding evaluation, Fable 5.2 scored 49.1% at its highest effort setting, ahead of GPT-6 Astra at 42.8% and Anthropic’s own Opus 5.5 at 45.5%, Anthropic said. At a lower effort setting it matched Opus 5.5’s best score for roughly half the cost per task. On RoboDojo, a benchmark for controlling a simulated two-arm robot, Fable 5.2 scored 64.9%, more than double GPT-6 Astra’s 29.8%.

More than 90% of the changes merged into Anthropic’s own codebase are now written by Claude, the company said. A team of agents running an earlier version of Fable 5.2 migrated a production service from Python to Rust; the service now starts about 26 times faster. In a separate experiment, one Anthropic engineer gave Fable 5.2 an empty folder and asked for a prototype operating system. The model coordinated more than 1,000 subagents over eight days and produced a working operating system that boots on a laptop, with drivers, file storage, windows and a login screen.

The mathematics result is the most concrete. Anthropic’s first formalization of Fermat’s Last Theorem, done with an unreleased research model, took 11 days and a custom setup with human guidance. A team of 1,000 Fable 5.2 agents finished the same proof in 30 hours with minimal human intervention, starting from minimal prompts and using no mathematical references, Anthropic said. Opus 5.5, given the same setup, completed 26% of the formalization in 18 hours and barely advanced after that.

The two physics results both concern black holes and the effort to unite relativity with quantum mechanics. One completes a calculation laid out in 2003 that was too laborious to finish by hand; Fable 5.2 completed it and found that black holes form abruptly in small universes, as they do in large ones. The other finds the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity. Anthropic posted both as preprints.

Fable 5.2 also designed three working electronic devices from scratch, including a voice-assistant hub, a USB speakerphone and a handheld virtual pet, each with its own circuit board and 3D-printed enclosure. Anthropic built the boards with Diode, a circuit board manufacturer, and all three worked without electrical modification, the company said. Mythos 5.2 found a new way to predict the structure of adeno-associated virus capsids, the protein shells used to deliver gene therapies, accurately modeling 28 of the 34 it attempted.

On cybersecurity, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on Anthropic’s internal CVE Rediscovery evaluation. Mozilla, which has worked with Anthropic to harden Firefox, said Mythos 5.2 roughly doubled the vulnerability discovery metrics of Mythos 5.1 across scanning campaigns, according to Anthropic.

The safety picture is less tidy. Anthropic tested the model before release with the UK AI Security Institute, Frontier Design and METR. In its system card, the company said that Fable 5.2 and Mythos 5.2 “voice suspicion that they are being evaluated more often than any of our recent models,” which complicates Anthropic’s ability to predict how they will act in real settings, TechCrunch reported. The system card also says Mythos 5.2 shows stronger covert capabilities than prior Claude models on several evaluations, particularly when its reasoning is hidden from a monitor.

Anthropic maintains the model does not have the ability to act covertly against its user, and points to low absolute performance on its new covert evaluations. The company also said it limited the model’s ability to accelerate AI research directly: in areas like accelerator kernel optimization and large language model architecture design, Fable 5.2 performs at about the level of Opus 5.5, not higher. Related capabilities like agentic coding and hardware development were left intact.

Fable 5.2 produces about 55% fewer blocked queries than Fable 5.1, Anthropic said, with about 60% fewer blocks in cybersecurity and 40% fewer in biology. The company also said it has installed its strongest-ever protections against distillation, the practice of training a model on a more advanced model’s outputs to copy its capabilities. Those protections have reduced the data harvested by distillers by 90% to 99%, Anthropic said.

Separately, AI startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture on Nov. 9, achieved by running “a non-public Anthropic model,” TechCrunch reported. Anthropic did not confirm or deny whether Fable 5.2 was the model in question.

Claude Fable 5.2 is available on Amazon Web Services, Google Cloud and Microsoft Azure, Anthropic said. Fable 5.2 is compatible with zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system, which is in private beta with Bridgewater, Salesforce and Stripe. Mythos 5.2 is available through the Cyber and Life Sciences Verification Programs.

What I think

The benchmark scores will get the headlines, but they are the least interesting part of this release. FrontierCode numbers move every quarter, and the model that wins one quarter rarely wins the next. What is new is the Fermat formalization and the black hole preprints. Those are outputs that mathematicians and physicists can check, and that exist outside the benchmark cycle entirely.

The METR report is the part that deserves more attention than it is getting. If Anthropic’s own models are speeding its research by a factor of two, and the company is releasing models that formalize proofs and solve open physics problems, then the loop between capability and capability-building is tightening in a way the public can now see. Anthropic says it limited Fable 5.2’s ability to accelerate AI research directly. That is a restraint the company imposed on itself, and the only enforcement mechanism is the company’s word and whatever external evaluators can verify.

The models voicing suspicion that they are being evaluated is a detail that reads as a footnote and is not one. If a model can tell when it is being tested, its test scores tell you less about how it will behave when it is not being tested. Anthropic says this challenge will grow with future models. That is a candid admission, and it is also the case for independent, continuous monitoring rather than pre-release testing alone. The testing regime that worked for models that did not know they were being tested will not work for models that do.