AI · 9 November 2026 · 4 min read
Fable 5.2 is on sale at $8 a million tokens. The black-hole problems are Anthropic’s own test
On 9 November Anthropic priced Claude Fable 5.2 at $8 per million input tokens and $40 per million output, and said answers come up to 30 percent faster than 5.1. Mythos 5.2, the less restricted name, is only for checked organisations. The physics, the proofs and the benchmark rows are the company’s.
Anthropic published Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. Fable is the one anyone can use. The page prices it at $8 per million input tokens and $40 per million output tokens. An optional fast mode, which Anthropic says can generate tokens up to 2.5 times quicker, is $16 and $80, double the ordinary rates. The same page says the models generate outputs up to 30 percent faster than Fable 5.1 and Mythos 5.1. Up to is the company’s word. TechCrunch reports the previous input price was $10. The page itself does not print the old price, so a 20 percent cut is a comparison, not a line in the release.
The demonstrations are also the company’s. Anthropic says that, in testing, the model solved two open problems in the physics of black holes, and that it formalised the proof of Fermat’s Last Theorem in 30 hours. It says a hypervisor the model built has more than 3,500 machine-checked proofs covering every case. None of those is a paper a reader can open from the announcement, and none is a result another lab has signed. On Anthropic’s own table, Fable 5.2 scores 89.3 percent on a partial offline set of OSWorld 2.1, against 83.5 percent for Fable 5.1, and 74.2 percent on Humanity’s Last Exam when tools are allowed, against 65.6 percent. On RoboDojo, a simulator, the same table puts it at 64.9 percent against 27.6 percent for 5.1. A simulator score is not a robot on a floor. The table is not an outside rerun, and it does not win every row: on AutomationBench, Anthropic’s own numbers put Gemini 4 Carbon at 53.9 percent and Fable 5.2 at 51.4.
Mythos is the name with the safeguards lifted, and it is not on general sale. Anthropic says Fable keeps strong cybersecurity and biology limits, which is why it can be offered to everyone. Organisations that want the wider model apply through a cyber check or a life-sciences check. The page also says the training was held back on the skills that would most directly speed up AI research. On accelerator kernels and on improving language-model architectures, Anthropic says both 5.2 models sit at about the level of Opus 5.5. A faster coding model, on the company’s own account, was not allowed to become a faster way of building the next model.