OpenAI Unveils Next-Gen o3 Model Set for Early Next Year Release

OpenAI has teased its next-generation frontier model, o3, with plans for a public release early next year, following the launch of o3-mini.

OpenAI has concluded its 12 Days of OpenAI livestream series with a sneak peek at its next-generation frontier model, dubbed o3. Sam Altman, OpenAI’s CEO, humorously noted the naming choice was influenced by a nod to Telefónica and OpenAI’s tradition of unconventional naming. The o3 model, while not yet ready for public use, will first be accessible to researchers for safety testing purposes.

Accompanying the announcement was the introduction of o3-mini, a model OpenAI plans to release by the end of January, with o3 to follow shortly thereafter. The o3 model boasts significant performance improvements over its predecessor, o1, demonstrated by its 96.7% accuracy on the American Invitational Mathematics Examination, a notable leap from o1’s 83.3%.

Mark Chen, OpenAI’s senior vice president of research, highlighted that o3’s performance was so impressive on standard benchmarks that the company sought more challenging tests, such as the ARC-AGI benchmark, which evaluates an AI’s ability to learn and intuit in real-time. The o3 model achieved scores of 75.7% and 87.5% on the ARC-AGI test under low and high compute settings, respectively, marking a significant milestone towards artificial general intelligence.

Additionally, OpenAI showcased o3-mini, which utilizes the Adaptive Thinking Time API to offer varying reasoning modes, allowing users to adjust the model’s problem-solving time. This innovation enables o3-mini to match the performance of the current o1 model at a reduced computational cost, with a public release anticipated before o3.