OpenAI claims gold-medal math performance on 2025 IMO

Tackling the International Math Olympiad (IMO) has long been considered one of AI's grandest reasoning hurdles. Many researchers believed gold-medal performance by a pure language model was still years away, but OpenAI just staked its claim.

OpenAI announced that its experimental general reasoning LLM achieved gold-level performance in an evaluation modeled directly after the 2025 IMO.

Human rules, zero external tools

The evaluation mirrored the exact conditions human prodigies face. The model had to tackle the exam across two 4.5-hour sessions, writing full natural language proofs from scratch.

There was no internet access and no external code or calculators running in the background. The model had to reason through pure text generation.

According to OpenAI, the unnamed system cracked 5 out of the 6 problems, racking up 35 out of 42 possible points. In the real competition, that score comfortably secures a gold medal.

To grade the complex proofs, OpenAI brought in three former IMO medalists. Each submission was evaluated independently, and points were awarded only after all three reached a unanimous consensus.

DeepMind pushes back

The claim immediately sparked pushback from rival Google DeepMind.

DeepMind publicly challenged the result, arguing that the IMO relies on confidential, internal marking guidelines. Without official grading under those rubrics, DeepMind stated that no legitimate gold claim can be made.

Competitive friction here makes sense. Solving the IMO has become the definitive proving ground for frontier reasoning models, and bragging rights carry enormous weight across the research community.

Crucially, OpenAI noted that this evaluation was run on an unreleased, experimental model, meaning the tech is still locked behind lab doors.

Why it matters

For everyday developers, you cannot test this model yet. But it provides a clear preview of where general reasoning is heading: AI that does not just regurgitate code snippets, but genuinely solves novel, abstract logical puzzles that baffle 99% of humans.

For the industry, this reignites the battle between DeepMind and OpenAI over formal versus natural language reasoning. While DeepMind has leaned heavily into formal verification systems, OpenAI wants to prove that large-scale general LLMs can climb the math mountain on their own.

We think having former medalists grade the papers by consensus was a smart defensive move, but self-conducted evaluations will always invite skepticism. Until OpenAI lets official organizers grade the model in real time or ships the checkpoint to the public, rivals will keep questioning whether the gold is genuine.