Sam Altman introduces OpenAI's new language model o3

OpenAI’s o3 model surpasses previous benchmarks, bringing us closer to AGI with its remarkable performance in complex tasks. The upcoming o3-Mini release promises even more advancements in AI’s problem-solving abilities.

Open AI promotes the o3 model as a step towards AGI. In ARC-AGI, a test designed to assess how efficiently an AI system can acquire new skills outside of the data on which it has been trained, o1 achieved a score of between 25 and 32 per cent out of 100 per cent. Here, 85 per cent is considered ‘human level’. According to Open AI, o3 already achieved 87.5 per cent of the points.

Greg Kamradt, President of the ARC Prize Foundation, which developed the test, also took part in the livestream, confirmed the result and congratulated Open AI on this milestone. Kamradt announced that he would work with Open AI in the future to develop new benchmarks.

As of today, external security researchers can apply to receive test access for the o3-Mini model.

The first model of the o3 family o3-Mini will be released at the end of January, announced Open AI's CEO Sam Altman in a livestream.

The company advertised the capabilities of the new model series as being able to programme exceptionally well and solve complex mathematical tasks. The o3 model is said to achieve a score of 96.7 per cent in the AIME 2024 mathematics test. On average, the model only gives the wrong answer once per test. In scientific questions at PhD level, o3 achieved 87.7 per cent in the GPQA Diamond test.