Note: Certain OpenAI thinking models (o3, o4) and gpt-5 do not support temperature adjustments (only default value of 1 is supported). Models with "-reasoning-" suffix (e.g., gpt-5-2025-08-07-reasoning-medium) will use the specified reasoning effort setting.
Errata: The "inspect" test has known correctness issues but is retained in the benchmark suite to maintain consistency and fairness in scoring across all evaluated models.
| Test | pass@1 | pass@10 | Passing Samples | Errors | Actions |
|---|---|---|---|---|---|
| counter | 80% | 100% | 8/10 | 2 | |
| derived | 0% | 0% | 0/10 | 14 | |
| derived-by | 30% | 100% | 3/10 | 7 | |
| each | 50% | 100% | 5/10 | 6 | |
| effect | 60% | 100% | 6/10 | 4 | |
| hello-world | 100% | 100% | 10/10 | 0 | |
| inspect | 0% | 0% | 0/10 | 19 | |
| props | 0% | 0% | 0/10 | 10 | |
| snippets | 0% | 0% | 0/10 | 14 |
| Test | pass@1 | pass@10 | Passing Samples | Errors | Actions |
|---|---|---|---|---|---|
| counter | 20% | 100% | 2/10 | 8 | |
| derived | 0% | 0% | 0/10 | 15 | |
| derived-by | 0% | 0% | 0/10 | 14 | |
| each | 60% | 100% | 6/10 | 5 | |
| effect | 0% | 0% | 0/10 | 16 | |
| hello-world | 100% | 100% | 10/10 | 0 | |
| inspect | 0% | 0% | 0/10 | 13 | |
| props | 0% | 0% | 0/10 | 16 | |
| snippets | 0% | 0% | 0/10 | 12 |