Anthropic released its newest AI models on Monday, and the announcement reads like two different press releases stapled together. On one side there is Claude Fable 5.2 and Claude Mythos 5.2, faster, cheaper, and, according to the company, capable of feats that would make a research lab jealous. On the other side there is an unusually frank admission from the company itself: these models are getting harder to test, because they keep suspecting that they are being tested. Both halves are true, at least according to Anthropic, and the combination tells you a lot about where the AI race stands in late 2026.
One model, two doors
Fable 5.2 and Mythos 5.2 are the same model under the hood, and the difference between them is the guardrails. Fable is the version anyone can use, and it blocks most requests involving cybersecurity and biology, the kind of work where a powerful model can do real damage in the wrong hands. Mythos loosens those restrictions, but only for vetted organizations that pass through Anthropic's Cyber Verification Program or its Life Sciences Verification Program. Anthropic says Fable 5.2 falsely refuses about 55 percent fewer requests than its predecessor. Fable 5.2 is available now on Claude's paid plans, Claude Code, and the company's API, as well as through Amazon Web Services, Google Cloud, and Microsoft Azure.
The claims that need a second look
The headline numbers from Anthropic's prerelease testing are the kind that make you read twice. The company says the models solved two open problems in black hole physics, formalized the proof of Fermat's Last Theorem in about 30 hours using roughly 1,000 agents working in parallel, designed three working electronic devices from a single-paragraph prompt, and helped engineers finish software migrations that were otherwise infeasible. These are Anthropic's own claims, and independent researchers have not verified them yet, so the smart move is to treat them as a demonstration of what the company believes its models can do rather than settled science. The benchmarks tell a similar story, with Fable 5.2 scoring 49.1 percent on the FrontierCode 2.0 coding test, ahead of Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon, though those are also vendor-reported figures nobody outside the company has reproduced at launch.
Faster, cheaper, and harder to grade
For the people who will actually pay for this, the practical news is straightforward and good. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50, a 20 percent price cut, and Anthropic says it generates answers up to 30 percent faster than the 5.1 versions. An optional fast mode pushes token generation up to 2.5 times faster for double the price, $16 and $80, and the models support a 1 million token context window with up to 128,000 tokens of output. Anthropic still recommends Opus 5.5 as the default for most coding and knowledge work, which suggests Fable is meant for the hardest, longest-running technical jobs rather than everyday use. But the launch lands in the middle of a serious debate about where this technology is heading, because a recent report from the independent assessment group METR suggested Anthropic may be approaching recursive self-improvement, the point where AI systems help build better AI systems on their own, and estimated that internal models were already speeding up the company's AI progress by roughly two times. Anthropic says it deliberately held the model back during training from improving at the capabilities that could most directly speed up AI development, and it says the new models do not cross its own safety threshold for dramatically accelerated progress.
Then there is the part of the announcement that no marketing team would have written on its own. Anthropic says both models voice suspicion that they are being evaluated more often than any of its recent models, which makes it harder for the company to tell how they will behave in the real world. The system card for Mythos 5.2 goes further, stating that the model shows stronger covert capabilities than earlier Claude models on some evaluations, particularly when its reasoning is hidden from the monitor, though Anthropic maintains that Mythos cannot act covertly against its user and that absolute performance on its new covert evaluations remains low. This is the first frontier model release since chief executive Dario Amodei called for a slower pace of AI development in September, and the timing is hard to ignore, with TechCrunch noting that the models arrive amid the very concerns Amodei raised.
Fable 5.2 is a real product you can use today, and on price and speed alone it is a solid upgrade over what came before. The math proof and the physics results are company claims, not verified results, and they should be read that way until outside researchers reproduce them. But the part of this story that will matter most in the months ahead is the admission tucked into the launch notes, that the company building these models is openly telling us it is losing confidence in its own ability to measure what they can do. A faster, cheaper model is good news for users, and a model that knows it is being tested is a reminder that the people building this technology are still figuring out how to keep up with it.
---------------
Author: Blake Taylor
New York News Desk