OpenAI drops o3 and o4-mini with full agentic tool access

OpenAI just launched o3 and o4-mini, the newest additions to its reasoning series, and they come with a massive upgrade: full, autonomous tool use built directly into how they think.

Instead of just predicting the next word or answering in a single shot, these models can now pause, formulate a plan, and agentically chain together web searches, Python code execution, visual analysis, and image generation, typically wrapping up in under a minute.

Thinking with images and running code on the fly

The headline technical leap here is that tools are no longer an afterthought. OpenAI trained both models using large-scale reinforcement learning to reason about when and how to deploy tools to solve a problem.

This is most apparent in visual reasoning. For the first time, images are woven directly into the model's chain of thought. If you upload a blurry textbook diagram or a tilted whiteboard snapshot, the model can actually manipulate the image, rotating, zooming in, or cropping it mid-thought to decipher tricky details.

OpenAI o3 sits at the top of the lineup as the frontier powerhouse. It set new state-of-the-art records on SWE-bench and Codeforces without needing custom scaffolds, and expert testers found it made 20% fewer major errors on hard real-world tasks compared to o1.

Meanwhile, o4-mini delivers ultra-fast reasoning at a fraction of the cost. When paired with a Python interpreter, o4-mini scored a staggering 99.5% pass@1 on the 2025 AIME math exam. Alongside the models, OpenAI also sneaked out Codex CLI, an experimental terminal-based agent that lets developers feed local screenshots and sketches directly into their code workflow.

Why it matters

We think this release marks a decisive transition from passive chatbots to genuine autonomous agents.

Until now, models either hallucinated when facts ran dry or relied on clunky prompt templates to trigger searches. Teaching reasoning models to know their own limits and spin up a Python script or multi-hop web query mid-thought makes them exponentially more dependable for everyday research, data science, and debugging.

For the broader AI landscape, it puts immense pressure on rivals like Google and Anthropic. The competition is no longer just about who has the cleverest pre-trained base model, but who can best harness reinforcement learning at inference time to manage complex, multi-tool workflows.

The trade-off is observability. As models gain the freedom to browse the web, write and execute code, and synthesize outputs behind a one-minute thinking screen, catching subtle hallucinated data embedded in a Python script's output will become increasingly tricky for regular users.