GPT-6.2 Sol scored 43.8% on Cognition's FrontierCode 2.0 leaderboard with Codex at Max effort, 0.4 percentage points behind GPT-6.2 Astra at the same settings. Sol averaged $1.29 per rollout, about 22% of Astra's cost.
Tasks it can assess: Changes that open-source maintainers would accept, including correctness, test quality, scope control, and code style.
Tasks beyond its scope: Code quality or cost in other tool environments, repositories, and reasoning settings.
Applicable model version: GPT-6.2 Sol as listed on the leaderboard. The model snapshot ID is unavailable.
Test environment or client: FrontierCode 2.0 with Codex.
Reasoning effort and parameters: All three runs in the table use Max effort. Other inference settings are unavailable.
FrontierCode 2.0 contains 132 tasks. Open-source maintainers write the tasks and define what makes a change mergeable. Cognition grades the results with unit tests, rubrics, and verifiers. The leaderboard score is a weighted aggregate of rubric items. A solution that fails a blocking criterion scores 0 for that rollout. Rollout costs use public API rates, including public discounts.
Sol Max ranked 6th in the snapshot. Sol and Astra both had a 48.2% pass rate at Codex Max. Sol 6.2 scored 4.2 percentage points higher than Sol 6.1 and cost $0.26 more per rollout.
| Rank | Model (Max) | Harness | Score | Pass rate | Flag rate | Cost per rollout | Output tokens |
|---|---|---|---|---|---|---|---|
| 4 | GPT-6.2 Astra | Codex | 44.2% | 48.2% | 0.0% | $5.97 | 46.0k |
| 6 | GPT-6.2 Sol | Codex | 43.8% | 48.2% | 0.0% | $1.29 | 38.1k |
| 9 | GPT-6.1 Sol | Codex | 39.6% | 43.0% | 0.0% | $1.03 | 41.5k |
The score and pass rate are separate metrics on the leaderboard. The $1.29 estimate is the average rollout cost on this task set. The benchmark and model lineup can change, so the ranks reflect the collection date.
Open the FrontierCode 2.0 leaderboard and wait for the table to load. Check the model rows under Codex at Max. An independent rerun requires the same tasks, scoring rules, and run settings. The leaderboard does not publish complete per-task records for these three runs.
GPT-6.2 Sol