OpenAI launches o3 and o4-mini

Two new models have been added to OpenAI's o-series: o3 and o4-mini. Two visual reasoning models. Yes, the two models are new, not to be confused with 4o and 4o-mini or o3-mini. The latter is already available. o3 was announced by OpenAI CEO Sam Altman back in December. After all, he also announced that he wanted to give the models better names in the future.

The special thing about the o series is that it is supposed to be particularly intelligent. The new models can access a whole range of tools. According to OpenAI's blog post, this includes web browsing, Python, image generation, image and file analysis, canvas, searching through files and the memory function. The models are also particularly good at solving complex mathematical problems, coding and scientific tasks – including visual strength. The models decide for themselves when they need which skills. This is another important step towards autonomous AI action.

As we know, the o models are trained using reinforcement learning via chains of thoughts. This refers to chains of thought in which confirmation is repeatedly sought for correct decisions. According to OpenAI, the models can now also think even better about their security guidelines, which makes them particularly robust. For example, they could understand and block attack scenarios. In an evaluation by OpenAI's security team, the Prepardeness Framework in an updated version, o3 and o4-mini performed particularly well. In none of three categories (biological and chemical capabilities, cybersecurity and AI self-improvement) was the security risk rated as high, says OpenAI. OpenAI has also published various benchmarks.

More performance and better image processing

OpenAI also describes how “more computing power produces better performance”. “By tracing the scaling path – back into RL – this time, we increased both the training computation and inference time by another order of magnitude and still saw significant performance gains.” This confirms that the performance of the models continues to improve the more they are allowed to think. At the same time, the new models are more cost-efficient than their predecessors.

In a separate blog post, OpenAI also explains how o3 and o4-mini can “think with images”. This means: “They not only see images, but can also integrate visual information directly into their reasoning chain.” This also means, for example, that the models can independently rotate images or zoom in on them and process these steps in the “thought process”.

OpenAI o3, o4-mini and o4-mini-high are now available for Plus, Pro and Team users. They replace o1, o3-mini and o3-mini-high. They can also be used via the API (Chat Completions + Responses API).

(emw)