Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresOpenAI's internal next iteration of GDPval, reported in the GPT-6.2 launch post: multistep assignments drawn from real consulting, finance and design projects, given the files and messages the team had when the work was assigned. The score's definition, grading and task count are not published, and OpenAI ran every model, including the Claude and Gemini comparators. Diagnostic.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6.2 Astra | max | 50.4 | |
| 2 | Gemini 4 Carbon | high | 48.8 | |
| 3 | GPT-6.2 Sol | max | 46.9 | |
| 4 | Claude Fable 5.2 | max | 44.8 | |
| 5 | GPT-6 Astra | max | 44.2 | |
| 6 | Claude Opus 5.5 | max | 43.5 | |
| 7 | GPT-6.1 Sol | max | 41.5 |