Claude Fable 5.2
Anthropic's flagship for the hardest open-ended work: 30% faster, 20% cheaper.
Claude Fable 5.2 is Anthropic's flagship model for the hardest open-ended work: complex coding, science, and robotics. Faster output and lower prices.
Try Claude Fable 5.2
Three upgrades separate this release from every model that came before. Each one removes a limit that used to mean a project simply could not be finished.
Claude Fable 5.2 spins up over a thousand sub-agents and keeps them coordinated. It built a bootable operating system this way in 8 days.
A one-million-token window holds an entire codebase or research thread at once, and the model reasons across the whole of it without loss.
Claude Fable 5.2 tops GPT-6 Astra and Gemini 4 Carbon on FrontierCode, chart reading, and robot control, setting the pace across the board.
Claude Fable 5.2 leads GPT-6 Astra and Gemini 4 Carbon on the hardest benchmarks, but the real difference is what it sustains: multi-day, thousand-agent projects rather than single answers. The table lines the flagships up on the capabilities that decide long-horizon work.
| Feature | Claude Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 42.8% | 41.1% |
| Chartography | 73.4% | Not reported | Not reported |
| RoboDojo | 64.9% | ~32% | Not reported |
| Sub-agent orchestration | 1,000+ agents, self-reviewed | Not supported | Not supported |
| Longest demonstrated run | 8 days, bootable OS prototype | Hours-scale | Hours-scale |
Write the goal in a brief, paste the relevant files, and name the constraints. Claude Fable 5.2 plans the whole approach before it starts.
It spins up sub-agents, assigns each one a task, and runs them while you watch. It reviews its own output and iterates until the job ships.
Check the finished deliverable, read the reasoning trail, and hand the rest to your team. Every step stays fully logged and ready to audit.
Hand it the project that has stalled your usual model, and Claude Fable 5.2 keeps working until it ships. Free to try online, no API key.
Try Claude Fable 5.2
Marcus Webb
Staff Engineer
“Our legacy Python service needed a rewrite nobody wanted to own, and Claude Fable 5.2 ran the migration to Rust with a team of sub-agents. It kept the plan coherent across days, caught its own errors, and handed back code that shipped. What we had shelved for a year as too expensive suddenly became a week of supervised work.”
Elena Petrova
Research Scientist
“I gave Claude Fable 5.2 an open question with a stack of papers, and it read all of them, flagged the one contradictory dataset, and returned a formal proof sketch. The long context held the whole thread at once, which no model before it managed. For serious research, this is a different category of tool.”
James Chen
Quant Developer
“We tested Claude Fable 5.2 on four open research problems, and it beat our best prior result on three of them while proposing a new angle we had missed. The thousand-agent orchestration is real, not marketing: it split each problem into sub-tasks, ran them in parallel, and stitched the results back coherently. It feels like a junior team.”
Sofia Marino
Robotics Engineer
“The RoboDojo numbers are what drew us in, and the real-world control is what kept us. Claude Fable 5.2 turned camera input into arm movements on tasks it had never been trained on, which usually requires months of specialist tuning. For a small lab with no robotics team, that is capability we simply could not build ourselves.”
Aisha Bello
Technical Lead
“We kept our daily work on Opus 5.5 and escalated only the hard parts to Claude Fable 5.2, exactly as Anthropic suggests. The split works: fast answers stay cheap, and the gnarly refactors that used to stall for weeks now get finished. Watching it run a multi-day migration was the first time a flagship earned its cost.”
Ryan Park
PhD Student
“Claude Fable 5.2 helped me formalize part of my thesis in Lean by coordinating sub-agents through the proof steps I had been stuck on for months. The million-token context kept my whole notation loaded at once, so nothing got dropped between sessions. I still do the thinking; the model removes the step-checking drudgery.”
