His reasoning models finally do what until now was impossible for them

For more than half a year, users of Chatgpt – and also those of the API – we have access to two types of models: the GPT models, such as GPT-4Oand reasoning -oriented models, such as O1. The problem is that, until now, we were forced to alternate between them for tasks that require complex thinking, web navigation or interpretation of images. As of today, that begins to change.

New models on stage. O3 and O4-mini models, presented this Wednesdayshorten distances with GPT models. For the first time, reasoning -centered models access tools that, so far, were exclusive to classic models. Namely:

• Analyze images (and think with them).
• Analyze files.
• Generate images.
• Navigate the web.
• Do research.
• Use advanced voice mode.

The keys to “visual reasoning”. The interesting thing is not that O3 read what is in a photo. It decides how to look at her. He knows if he has to turn it, expand a detail or ignore the irrelevant. That process is now part of its reasoning chain. He does not describe an image, but “think” about it to give us a better answer.

A remarkable jump. We are facing a series of improvements that the most demanding users will undoubtedly know how to appreciate. The reasoning models, it should be remembered, “think” before responding. They generate an internal chain of thought before offering us an answer.

They are not the best option for those looking for texts with literary dyes or quick responses on any subject. But they are ideal to code, reason scientifically and plan complex workflows, especially in environments where several steps are involved and Agents.

Taking this into account, and as expected, Openai has also improved the main capabilities of these models, making them more capable and precise.

• OpenAI O3: This model achieves a prominent performance in Swe-Bench verified (without customization), a test that measures coding skills, with a score of 69.1%. According to OpenAI, in evaluations carried out by external experts, O3 makes 20% less important errors that OPENAI O1 in difficult real -world tasks.
• OPENAI O4-MINI: It offers similar yield, with a score of 68.1%. To put it in context: O3-mini obtained 49.3% in the same test, while Claude 3.7 Sonnet reached 62.3%. This model is optimized for rapid and profitable reasoning, and yields especially well in mathematical, coding and visual tasks.

The model that was not going to launch. OpenAi has changed their minds. In February, Sam Altman ruled out launch O3 as an independent product. But just a few weeks ago, toHe gave a “change of plans”. This turn materializes today with the arrival of O3 and O4-mini, the new models that mark a new chapter in the company’s strategy.

On the way to chatgpt. From today, users of Chatgpt plusPro and Team can already start using O3 and O4-mini. In the coming weeks, O3-PRO will arrive, a more powerful version of the reasoning model, which will be available for subscribers of the Pro Plan. Meanwhile, these users can continue working with O1-PRO.

Images | OpenAI

In Xataka | Openai’s hypothetical social network does not want to connect people. Want your data to train your AI