OpenAI Wows the Crowd as New Scaling Law Passes Its First Test

OpenAI’s next-generation reasoning model—called o3 for a funny reason—seriously impressed researchers and developers last week.

Specifically, o3 and o3-mini (a smaller version of the model) scored impressively on a number of extremely difficult math and coding benchmarks, beating out competitors like Anthropic and Google by a lot. It also reached human levels of performance on ARC-AGI, a benchmark meant to test how AI models handle tasks that humans excel at such as complex problem-solving and pattern recognition.

This is great news for researchers worried about the slowing improvements they’ve been seeing from pretraining, the process of initially training models on tons of data to help them make sense of the world and the relationships between different concepts. (The idea that AI models improve the more data and compute you give them during pretraining is known as the scaling laws.) 

Reasoning models, as we noted six weeks ago, were meant to make up for the pretraining slowdown, by having models spend more computing resources “thinking,” or processing customer queries.