OpenAI today announced an improved version of its most capable yet artificial intelligence model so far—one that took even longer to consider the questions—just one day later Google announced the first model of its kind.
OpenAI's new model, called o3, replaces the company's o1 introduced in September. Like o1, the new model takes time to reflect on a problem to come up with better answers to questions that require step-by-step logical reasoning. (OpenAI chose to omit the “o2” moniker because it was already the name of a UK mobile provider.)
“We see this as the beginning of the next phase of AI,” OpenAI CEO Sam Altman said during Friday's livestream. “Where you can use these models to do increasingly complex tasks that require a lot of inference.”
The o3 model scored much higher on several metrics than its predecessor, OpenAI said, including those measuring complex coding-related skills as well as mathematical prowess and advanced science. It is three times better than o1 at answering do questions ARC-AGIa benchmark designed to test the reasoning ability of AI models on extremely difficult logic and math problems they encounter for the first time.
Google is pursuing a similar line of research. Noam Shazeer, a Google researcher, yesterday revealed in a post on X that the company has developed its own reasoning model, called Gemini 2.0 Flash Thinking. Google CEO Sundar Pichai called it “our most thoughtful model yet” in his post. separate article.
The two dueling models show that the competition between OpenAI and Google is fiercer than ever. It's important for OpenAI to demonstrate that it can continue to make progress as it seeks to attract more investment and build a profitable business. Meanwhile, Google is trying to show that it is still at the forefront of AI research.
The new models also show that AI companies are increasingly interested in simply scaling AI models to gain greater intelligence from them.
OpenAI says there are two versions of the new model, o3 and o3-mini. The company has not yet made the models public but said it will invite outsiders to sign up to test them. OpenAI today also revealed more details about the techniques used to align o1. This involves having sample reasons about the nature of the request being made to interrogate whether it violates its protective barriers.