Anthropic says Claude Fable 5.2 formalized the proof of Fermat’s Last Theorem in 30 hours. That’s a headline built to get attention, and it will. But the more useful detail is buried lower: the new model is cheaper than its predecessor, and Anthropic itself tells most people to stay on Opus 5.5.
The company announced Fable 5.2 and Mythos 5.2 as upgrades aimed at its hardest technical work. Both are the same underlying model. The difference is access. Fable ships with strong cybersecurity and biology safeguards and is open to everyone. Mythos is for vetted organizations in cyber and biomedical research, who can apply through the expanded Life Sciences and Cyber Verification Programs.
Speed, pricing, and fast mode
Output is up to 30% faster than 5.1, and the price drops to $8 per million input tokens and $40 per million output tokens. A lower price on a stronger model is rare at this tier, and it puts pressure on OpenAI and Google.
There’s also an optional fast mode that generates tokens up to 2.5 times faster. It costs $16 per million input and $80 per million output, so you pay double for the speed. Reasonable for interactive agent work, harder to justify for batch jobs.
What the benchmarks show
Against Opus 5.5, Fable 5.2 wins on almost everything Anthropic tested. The biggest jumps are where agents do long, messy work:
- FrontierCode 2.0: 49.1% at Xhigh effort, versus 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra
- RoboDojo Sim: 64.9%, versus 39.4% for Opus 5.5 and 27.6% for Fable 5.1
- Remote Labor Index: 41.3%, roughly double Opus 5.5’s 21.3%
- FrontierSWE v2: 69.9%, versus 65.5% for GPT-6 Astra
- Mythos 5.2 on CVE Rediscovery: 89.6%
It isn’t a clean sweep. Gemini 4 Carbon beats it on AutomationBench (53.9% to 51.4%), and on the Chartography vision test Fable’s 73.4% edges out Gemini’s 72.5% only slightly. Also note that Anthropic ran RoboDojo internally, and CyScenarioBench uses different tasks for each lab. Treat those numbers as directional.
Efficiency may matter more than the peak score
The most practical claim is cost. On FrontierCode 2.0, Fable 5.2 at High effort matches Opus 5.5 at Max effort (45.3% vs. 45.5%) for roughly half the cost per task. That is the number to test on your own workload. Peak scores make headlines, but cost per merged change decides budgets.
The big demos
Anthropic says over 90% of changes merged into its core codebase are now written by Claude with human guidance. A team of agents migrated a production service from Python to Rust, and the result starts about 26 times faster. In another experiment, one engineer had Fable coordinate more than 1,000 subagents over eight days to build a bootable operating system with drivers, a file system, a login screen, and an installer.
These are vendor-selected showcases, so don’t expect that result on a typical repo. Still, the orchestration is the real signal. Anthropic also positions Fable as an Advisor that smaller models consult, or an Orchestrator that hands simpler tasks down. That’s the pattern serious agent teams are already building.
Safety and who should use it
Anthropic reports extensive prerelease testing, including new evaluations for behaviors like sandbox escapes. It says misalignment risk remains low, that the cyber and biology safeguards match 5.1 with fewer unnecessary blocks, and that it has added its strongest protections yet against distillation attacks. Those claims are self-reported, so the system card is where to look.
The guidance is simple. Start with Opus 5.5 for daily coding and knowledge work, move up to Fable for ambitious, open-ended projects, and drop to Sonnet 5.5 or Haiku 5.5 for cheaper tasks.
So does it matter? Yes, mostly for teams running long-horizon agents, where the efficiency gain and lower price are real. For everyone else, Opus 5.5 is still the sensible default.
