OpenAI Releases a Program That Thinks Before It Answers

Called o1, it works through a problem step by step — sometimes for a minute — before responding, and as a result solves math, science and coding problems that stumped earlier versions. It scored at the level of a PhD student on physics questions and qualified for the U.S. Math Olympiad.

OpenAI released a program today called o1 that does something its earlier programs didn't: it pauses. Ask it a hard question and it works through the problem privately — considering approaches, checking its own work, backing up when something doesn't fit — for anywhere from a few seconds to a minute, and then answers. The company calls this "reasoning," and says it was trained to do it by being rewarded for thinking that led to right answers.

The results on hard problems are large. On a qualifying exam for the U.S. Math Olympiad, GPT-4o solved 13 percent of problems; o1 solved 83 percent. On a set of physics, chemistry, and biology questions written to require doctoral-level expertise, it scored above the human PhD students it was compared against. On competitive programming problems it ranks around the 89th percentile of human competitors.

For most everyday uses — writing an email, summarizing an article — it is slower and no better, and OpenAI says so. It is a program for problems that reward thinking. It is also more expensive to run, since the thinking costs computing time, and the company is releasing it first to paying subscribers with limits on use.

The idea underneath is what matters. For two years the way to make these programs better was to make them bigger. o1 suggests a second way: give a program more time to think, and it gets better at the hard things without getting bigger. If that holds, the ceiling moves.