OpenAI's reasoning system snags gold at IOI

OpenAI just went toe-to-toe with the world's brightest young coders, and it almost beat them all.

Competing in the online track of the 2025 International Olympiad in Informatics (IOI), an OpenAI reasoning system scored high enough to hit gold medal territory. Out of roughly 330 elite high school competitors, only five human participants outscored the model, putting it at sixth overall and first among all AI entrants.

Even wilder: it played strictly by the human rulebook, operating under the standard five-hour clock and capped at 50 submissions.

Ditching the complex scaffolding

Just one year ago, OpenAI's programming performance was fairly mediocre. In 2024, the team finished around the 49th percentile, missing out on a bronze medal entirely.

Back then, the lab relied on a heavily fine-tuned model wrapped in a massive, handcrafted scaffolding packed with synthetic test cases and hardcoded features. It was complex, clunky, and fragile.

This year, the researchers took the exact opposite approach. They stripped away the engineering bloat, deploying an ensemble of general-purpose models paired with a lightweight selection layer that simply picked the best output using another model and a basic heuristic. That minimalist tweak rocketed the system to the 98th percentile.

One brain for math and code

Perhaps the biggest takeaway is that OpenAI didn't even build a dedicated coding model for this tournament.

OpenAI researcher Noam Brown revealed that the core logic engine behind this IOI run is the exact same model that earned a gold medal at the International Mathematical Olympiad (IMO) weeks ago. After realizing the model's competitive math prowess translated seamlessly into software engineering, the team deployed it without any code-specific modifications.

Over the span of just a few weeks, OpenAI's reasoning stack has taken second place on AtCoder, hit gold level at the IMO, and secured gold level at the IOI. Brown noted that by next year, the raw model itself will likely outshine any custom scaffolding humans could build for it.

Why it matters

We think this run marks a major turning point in how frontier models will be built.

Until recently, the playbook for technical domains was heavy specialization: train a niche coding model, build elaborate agent frameworks, and handhold the AI through tricky edge cases. OpenAI just proved that raw, generalized reasoning can bulldoze complex domain-specific tasks without any custom tuning.

For developers, the writing is on the wall. The thousands of hours spent building complex wrappers, brittle test generators, and hardcoded prompt loops are depreciating fast. If generalized logic can effortlessly jump from formal math proofs to Olympic-level dynamic programming, domain-specific AI wrappers are going to face an existential test very soon.