OpenAI publishes paper on competitive programming with reasoning models

OpenAI published a paper comparing two approaches to competitive programming with its reasoning models. The first, a version of o1 fitted with hand-engineered, contest-specific strategies, competed live at the 2024 International Olympiad in Informatics and placed in the 49th percentile — reaching gold-medal level only under relaxed time and submission constraints, not the actual competition rules. The second, the more general-purpose o3, subsequently achieved a gold-medal result at IOI 2024 conditions without any domain-specific engineering, and separately reached a Codeforces rating the paper described as on par with elite human competitors.

The paper’s central argument was that scaling general-purpose reinforcement learning produced better competitive-programming performance than building narrow, hand-crafted strategies on top of a fixed model — a claim consistent with OpenAI’s broader bet that the same reasoning training that improved math and science performance would generalise to structured problem-solving domains without task-specific tooling.

The result added competitive programming to the list of benchmarks — following mathematics olympiad problems and PhD-level science questions — where OpenAI’s reasoning models had moved from clearly below top human performance to matching or exceeding it within about a year, a pace that fed into the broader argument over how much headroom remained in problems with clean, verifiable answers as opposed to open-ended real-world tasks.