April was the month reasoning models learned to use tools.
o3 and o4-mini
On April 16, OpenAI released o3 and o4-mini.¹ Like earlier reasoning models, they work through a problem before answering. The new part is that they can use tools while they reason. They can search the web, run Python, and look at an uploaded image, then use what they find in the next step of their thinking.
OpenAI calls o3 its most capable reasoning model so far. o4-mini is the cheaper, faster option: $1.10 per million input tokens, against $10 for o3.
Why Tool Use Matters
A model that reasons without tools can only rearrange what it already knows. A model that can look things up and run code can check its own work. Answers get better. They also get easier to audit, because the tool calls show how the model got to its answer.
Confidently Wrong
The risk is that longer reasoning makes wrong answers more convincing. Ten careful-looking steps built on one bad assumption still read as careful. For anything high-stakes, I want:
- The full trace logged: reasoning steps and tool calls stored next to the final answer
- A human check on the starting assumptions, since that is where a long chain usually goes wrong
- Evaluations on your own data, because public benchmarks won't show your failure modes
For Product Teams
Reasoning models are slower and cost more per call. Use them where the extra thinking pays off, like multi-step analysis, debugging, and planning. Keep a fast model for everything else. Deciding which requests go to which model is now a product decision, and it has a big effect on both cost and user experience.
Where are reasoning models earning their cost in your products, and where aren't they?