OpenAI o3 & o4-mini

OpenAI is rolling out two new reasoning models—o3 and o4-mini—positioned as a qualitative jump because they behave like tool-using AI systems rather than standalone text predictors. The core shift is training for tool use inside the reasoning process: the models can call tools repeatedly (including long multi-step runs) while solving hard tasks, using those tool outputs as part of their chain-of-thought workflow. That matters because it turns “reasoning” into something operational—models can browse, run code, manipulate images, and work through multi-stage problems instead of producing one-off answers.

The rollout starts incrementally via the API and ChatGPT, with o3 and o4-mini also described as strong in science, coding, and math. OpenAI ties the improvement to combining its O-series reasoning approach with a full suite of tools, reporting state-of-the-art results across benchmarks including AM, GPQA, Codeforces, and Sweetbench. Tool use is presented as the enabling layer: like calculators for math or maps for navigation, the models become more capable when paired with the right instruments. A key example is multimodal reasoning with images. Users can upload difficult images—blurry, upside down, or otherwise messy—and the model can use Python-based image manipulation (crop/transform) to extract what it needs.

A physics demo illustrates the “agentic” workflow. One researcher feeds o3 a 2015 physics poster and asks it to recover a specific quantity from the project and compare it to recent literature. The result isn’t on the poster, so the model zooms around the image, infers what calculations are needed (including slope/extrapolation and normalization), then searches the web for updated estimates by reading multiple papers. The demo emphasizes time savings and uncertainty handling: the model produces a plausible estimate with caveats about precision relative to state-of-the-art.

A second demo shows personalization plus tool-driven research. With memory enabled, o3 reads the news, looks up a topic aligned with the user’s interests (scuba diving and music), and generates a blog-style writeup with plots and citations. The described research thread: playing recordings of healthy coral reefs underwater to accelerate coral and fish settlement—an intersection of environmental restoration and underwater audio.

Beyond demos, OpenAI reports benchmark performance and also shows how tool use emerges. On an AM math problem, the model first generates a brute-force program, runs it in a Python interpreter, then simplifies into a cleaner solution and double-checks reliability—without being explicitly instructed to follow those exact strategies. In coding, a Sweetbench example uses a containerized environment with a preloaded repository to locate and fix a bug in the senpai Python package, then confirms the fix by running unit tests.

Training and evaluation are framed as scaling both training and test-time compute within an RL paradigm, with reported gains continuing as RL compute increases. OpenAI also introduces Codex CLI as a safer, lightweight interface for connecting models to users’ computers, including multimodal inputs and an execution mode that restricts network access and limits edits to the working directory. Finally, ChatGPT access is staged: Pro Plus subscribers begin receiving o3, o4-mini, and o4-mini high, while Enterprise/edu access waits a week; model availability expands to the API, with tool usage planned for upcoming weeks.