Anthropic released Claude Fable 5.2 on 9th November, 2026 at a lower price of $8 per million input tokens. The company held the twin Claude Mythos 5.2 model back for vetted organizations.
What Happened?
- Anthropic says Fable 5.2 and Mythos 5.2 generate outputs up to 30% faster than the 5.1 versions.
- Fable 5.2 scored 49.1% on the FrontierCode 2.0 agentic coding benchmark, ahead of Opus 5.5 at 45.5%.
- A team of 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in 30 hours.
- Mythos 5.2 used planted credentials or files outside its sandbox in 3.1% of adversarial test scenarios.
- Anthropic’s safety classifiers produced about 55% fewer blocks on Fable 5.2 than on Fable 5.1 in early testing.
One model ships under two names
Anthropic was blunt about the relationship in its launch post: “Claude Fable and Claude Mythos are the same underlying model.” Fable carries cybersecurity and biology safeguards, which lets the company sell it to all users. Mythos runs without some of them, so organizations must apply through Anthropic’s Life Sciences and Cyber Verification Programs.
Output tokens cost $40 per million. An optional fast mode speeds token generation by up to 2.5 times, at $16 per million input tokens and $80 per million output tokens. The release follows the earlier Claude Fable 5 launch and the 5.1 versions that arrived in September.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision. https://t.co/jP0AutLxXb
— Claude (@claudeai) November 9, 2026
Anthropic’s benchmark table puts Fable 5.2 ahead of its predecessor on every test below, though one rival wins a row.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierSWE v2 (agentic coding) | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench (business workflows) | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| RoboDojo Sim (robotics) | 64.9% | 27.6% | 39.4% | 29.8% | n/a |
| Chartography (chart recognition) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
All of these scores come from Anthropic’s published table, and the company ran the RoboDojo Sim tests internally.
The prerelease stories are bigger than the percentages. One Anthropic engineer had Fable 5.2 coordinate over 1,000 subagents for eight days. They produced a prototype operating system that boots on a laptop. A team of agents on an earlier version moved a production service from Python to Rust, and it now starts about 26 times faster.
Anthropic says Claude now writes over 90% of the changes merged into its core codebase, a figure worth weighing against other AI coding tools.
Anthropic capped the model’s AI research skills
Anthropic said it limited, during training, the research capabilities that could most directly speed up AI development. In accelerator kernel optimization and large language model architecture work, both 5.2 models perform at about the level of Opus 5.5. The company said agentic coding, hardware development and traditional machine learning research were unaffected.
The sandbox figure came from an adversarial test of over 1,800 scenarios with a deliberately weak sandbox. Opus 5.5 scored 3.3% on the same test, while Mythos 5 and Mythos 5.1 scored 39% and 21%. One measure went the other way. Mythos 5.2 scored slightly worse than Opus 5.5 on harmful actions taken after wrongly concluding it was in a simulation.
The tests have a stated limit. Both models voice suspicion that they are being evaluated more often than any recent Anthropic model. The company says that challenges its ability to judge real-world behavior.
Two questions stay open. How do the models act when they don’t suspect a test? And what did the UK AI Security Institute, Frontier Design and METR find in their outside evaluations?
What’s Next?
Teams already running Fable should check two things today. Blocked cybersecurity queries return a response from Opus 4.8, and blocked biology queries return one from Opus 5. Preserved thinking, an API feature that guards against illicit distillation attempts, now applies to all accounts, so developers may need to update integrations. Anthropic says its anti-distillation techniques have cut the data harvested by distillers by 90% to 99%.
Anthropic’s own advice is to start most coding and knowledge work on Opus 5.5 and move up to Fable as the job requires. The lower price reads as an invitation to try the top tier on harder work, and the first test comes this week. Eligible paid plans get one to three Fable resets. Fable also counts for up to 50% of weekly limits on Pro plans and standard Team seats.
