ICPC World Finals 2025 result (GPT-5 plus experimental model)
OpenAI's reasoning systems solve 12 of 12 ICPC World Finals problems, ahead of the best human team (11 of 12).
An ensemble of general-purpose models. GPT-5 solved 11 problems and an experimental reasoning model solved the hardest one on its 9th submission and chose which answers to submit. No contest-specific training. Models had a code-execution sandbox but no internet. Google DeepMind's Gemini 2.5 Deep Think solved 10.
- Date
- Wednesday, 17 September 2025
- Lab
- OpenAI
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| ICPC problems solved | 12 of 12 11 correct on first answer; hardest on 9th submission; best human team 11/12 | company |
| Gemini 2.5 Deep Think, same contest | 10 of 12 for comparison; reported by Google DeepMind | company |
Company-reported (via researcher statements quoted by Simon Willison, with a 2025-09-27 clarification that a code sandbox was used); OpenAI's own post was not opened.
Sources
This record was checked and corrected against its sources on 6 October 2026. How we check