[OpenAI GPT-5] ICPC World Finals Achievement
OpenAI GPT-5 and an internal experimental reasoning model solved all 12 problems at the 2025 ICPC World Finals, producing top scores consistent with a first-place human ranking. GPT-5 produced correct solutions for 11 of the 12 problems on first submission; the final, most challenging problem was completed after collaboration with an experimental model following a combined nine submissions. The team used a simple approach—simultaneously generating multiple candidate solutions and using the experimental model to select the best submission—running under the same time, memory, and sandboxed constraints as human competitors. The results underscore recent advances in general-purpose reasoning, competitive performance across math and coding contests, and product-level options allowing users to tune GPT-5’s thinking time.
Key insights
GPT-5 and experimental model achieved full problem set completion
The models solved all 12 ICPC World Finals problems, producing results equivalent to a first-place human ranking and demonstrating top-tier contest performance.
Detailed submission performance
GPT-5 solved 11 of 12 problems on the first submission to the official judging environment; the final problem was solved after the experimental reasoning model assisted, with a combined nine submissions reported.
Operational method: generate-and-select pipeline
A straightforward pipeline generated multiple candidate solutions from GPT-5 and the experimental model in parallel, then used the experimental model to select which solutions to submit—no complex scaffolding was used.
Models evaluated under standard contest constraints
All models competed under identical hidden test cases, time limits, memory constraints, hardware specifications, and sandboxed evaluation environments applied to human participants.
Product tuning: user-selectable thinking time for GPT-5
Users of Plus, Pro, and Business can now select GPT-5 thinking-time modes (Standard or Extended; Pro also offers Light and Heavy) to trade off speed and depth of reasoning in the chat composer.