GPT-6 Astra leads overall but trails Claude on coding

Epoch AI broke its Capabilities Index down by domain, and the split is worth reading before you pick a model for a specific job. GPT-6 Astra takes the top overall score at 166, ahead of Claude Fable 5.1 at 164 and OpenAI's own GPT-5.6 Sol at 162. On math the gap is wide: Astra's Math-ECI hits 169.83, a record across every model Epoch has tested. Software engineering tells a different story. There Astra scores 163.58 on the SWE-ECI, behind Claude Fable 5.1's 167.44. The single best model overall is not the best model for coding.

The numbers come with real caveats, and Epoch is direct about them. Each domain score rests on only three to six benchmark results, which leaves wide confidence intervals, and the estimates move as new results arrive. Astra's software engineering score dropped 2.9 points in ten days as more coding evaluations came in. So the ranking is a snapshot, not a settled verdict, but the shape of it is consistent: strength in one domain does not transfer cleanly to another, even within a single frontier model.

Why it matters

If you choose models by a single leaderboard number, you may be paying for math ability you do not use while leaving coding performance on the table. Match the model to the domain you actually work in, and treat any headline capability score as a starting point to test against your own tasks.