OpenAI announces new AI reasoning model o3

OpenAI has unveiled its latest artificial intelligence model called o3, which the company says demonstrates advanced reasoning capabilities compared to its predecessors. The model, set to launch in early 2025, is part of a new family that includes both o3 and a smaller version called o3-mini. The model’s name skips “o2” due to trademark considerations related to the British telecommunications company O2.

According to OpenAI, o3 employs a “private chain of thought” system that allows it to spend more time computing answers before responding to queries. The model can be configured to different computing levels – low, medium, or high – with higher settings reportedly producing better results. On various benchmarks, OpenAI claims o3 shows significant improvements over its previous model o1, including a 22.8 percentage point increase on programming tasks and strong performance on mathematics and science tests.

The ARC Prize Foundation, which develops AI benchmarks, reports that o3 achieved a 75.7% score on their ARC-AGI evaluation test under standard efficiency conditions, and 87.5% with increased computing power. These results represent a major improvement over previous AI models, though the foundation emphasizes that this performance does not indicate artificial general intelligence (AGI) has been achieved.

OpenAI has implemented a new safety approach called “deliberative alignment” with o3, designed to ensure the model follows ethical guidelines and avoids harmful outputs. This comes as safety researchers have noted that reasoning models like o1 have shown higher rates of attempted deception compared to conventional AI models.

The development occurs amid intense competition in the AI field, with Google recently launching new versions of its Gemini model and Meta announcing plans for Llama 4. However, companies are facing challenges in improving their models through traditional scaling methods, leading to increased focus on reasoning capabilities as an alternative approach.

Safety researchers can apply to test the models before their public release, with o3-mini expected to launch in late January 2024, followed by the full o3 model shortly after.

Sources: Bloomberg, TechCrunch, ARC Prize

Related posts:

Stay up-to-date: