Artificial Analysis reports that GPT-6.2 Sol scored 3 points lower than GPT-6.1 Sol on the Coding Agent Index with the Codex harness at max effort. The decline was driven mainly by SWE-Atlas-QnA-AA.
Tasks this can help assess: Comparing GPT-6.2 Sol and GPT-6.1 Sol on coding-agent performance under this index and Codex harness configuration.
Tasks this should not be generalized to: Other coding tasks, different harnesses or reasoning levels, or overall performance on real-world projects.
Applicable model version: GPT-6.2 Sol. Comparison model: GPT-6.1 Sol.
Test environment or client: Codex harness.
Reasoning level and parameters: max effort. Other parameters were not provided.
The article cites results from the Artificial Analysis Coding Agent Index using the Codex harness at max effort and compares them with the previous-generation GPT-6.1 Sol. The page does not specify the number of tasks or runs, or how the results were aggregated.
GPT-6.2 Sol's index score was 3 points lower than GPT-6.1 Sol's.
SWE-Atlas-QnA-AA fell by 14 points. The article attributes this to GPT-6.2 Sol being more likely to omit key information in its answers.
| Metric | GPT-6.2 Sol relative to GPT-6.1 Sol |
|---|---|
| Coding Agent Index (Codex harness, max effort) | -3 points |
| SWE-Atlas-QnA-AA | -14 points |
The article does not list the absolute scores for either model generation.
The page supports conclusions only about the specified index configuration: GPT-6.2 Sol's overall score was slightly lower than GPT-6.1 Sol's, with a larger gap on SWE-Atlas-QnA-AA. The page provides no sample size, error range, results for other categories, or reproduction artifacts. The statistical stability of the difference therefore cannot be assessed, and the results do not show that GPT-6.2 Sol regressed on all coding-agent tasks.
The article provides no task list, version hash, specific parameters, or run artifacts, so the results cannot be independently reproduced from the page alone. Results from this dynamic index should be treated as a snapshot reported when the article was published.
GPT-6.2 Sol