What are OpenAI o3 and o4-mini, and what changes when a reasoning model can use tools?

Yesterday, April 16, 2025, OpenAI released two new models in its "o-series," the family trained to think before it answers: o3 and o4-mini. Paying ChatGPT users can pick them today, developers can call them through the API, and free users can try o4-mini.

The headline is not only that they score higher on tests. It is that these reasoning models can now use every tool ChatGPT offers, and choose for themselves when to use them. Ask a question that needs current numbers, a calculation and a chart, and the model can search, compute and draw in one go, thinking between each step.

What OpenAI released

OpenAI calls o3 its most powerful reasoning model, strong at coding, math, science and understanding images. In evaluations by outside experts, it made 20 percent fewer major errors than o1, the first model in the series, on difficult real-world tasks. o4-mini is smaller and tuned for speed and cost. It does especially well in math, coding and visual tasks, and because it is cheaper to run, ChatGPT allows many more messages with it.

The scores are OpenAI's, as reported by TechCrunch, which notes that o3-mini scored 49.3% on the same coding test. A token is a piece of a word; a million tokens is roughly 750,000 words.

In ChatGPT, the new models replace o1, o3-mini and o3-mini-high in the model menu, alongside a variant called o4-mini-high that spends more effort on each answer. Enterprise and Edu customers get access a week later. OpenAI also says a more powerful o3-pro, with full tool support, should follow in a few weeks.

How a reasoning model with tools works

A reasoning model writes a private chain of thought, a series of intermediate steps, before giving its final answer. You can picture a student working on scrap paper before writing a clean answer on the test. Earlier o-series models in ChatGPT did that thinking largely on their own. o3 and o4-mini can pause in the middle of their thinking to use a tool, read the result, and keep reasoning.

OpenAI trained this behavior with reinforcement learning, a method where the model tries many approaches and is rewarded for the ones that work. The company says it taught the models not just how to use tools but when they are worth using. It also reports that reinforcement learning follows the same pattern seen when training earlier GPT models: more computing power gives better results, and so does letting the model think longer.

Thinking with images is the second new ability. Earlier models could describe a photo. These models can bring an image into the reasoning itself, and use tools to zoom in, rotate or crop it as they work. OpenAI says that helps with blurry, upside-down or low-quality pictures, such as a quick phone photo of a whiteboard.

A worked example

OpenAI's announcement offers this kind of request: how will summer energy use in California compare with last year? Here is how the model can break it down:

1. Search the web for public utility data on past and current use.
2. Run Python to clean the numbers and build a simple forecast.
3. Make a chart comparing the two summers.
4. Explain the main factors behind the prediction, and search again if something is missing.

You can write a similar prompt yourself:

I have attached a photo of my handwritten budget for April.
Read the amounts, add them up by category, and make a bar chart.
Then tell me which category grew the most compared with March,
which is in the second photo. If any number is unclear, say so.

The last sentence matters: it gives the model permission to flag what it cannot read rather than guess.

Why it matters

Tool use turns a reasoning model from a careful thinker into something closer to a research assistant. TechCrunch notes that reasoning models now lead the field, as several labs, including Google, Anthropic and DeepSeek, release their own. OpenAI says o3 and o4-mini are usually both smarter and cheaper than the models they replace, which makes careful reasoning practical for everyday questions.

OpenAI also describes the release as part of a direction: bringing the reasoning of the o-series together with the natural conversation and tool use of its GPT models. On safety, the company says it rebuilt its refusal training data for these models and that its tests placed both below the "High" risk level of its Preparedness Framework, the internal scale it uses to judge dangerous capabilities.

The same announcement introduced Codex CLI, an open-source coding agent for the terminal that uses these models by default.

Which should you pick in ChatGPT, o3 or o4-mini?

Choose o3 for questions where quality matters more than speed: a complex analysis, a tricky piece of code, a chart or diagram you need interpreted carefully, or research that combines several sources. It usually answers in under a minute, but it uses more of your message allowance.

Choose o4-mini for the many everyday questions that still benefit from step-by-step thinking, such as math homework, a quick data summary or a short coding task. Its higher limits make it the better default for frequent use, and o4-mini-high is there when you want a little more care.

Whichever you pick, give the model the material it needs, say what form the answer should take, and ask it to point out anything it could not verify. Reasoning reduces mistakes; it does not remove them.