one year on
OpenAI launches o3 and o4-mini reasoning models with tool use and image thinking
The new models can browse the web, run Python, and process images during chain-of-thought reasoning, marking a shift toward integrated tool use inside AI reasoning.
OpenAI today released o3 and o4-mini, a new pair of reasoning models that pause to work through questions before answering. The company calls o3 its most advanced general reasoning model yet, outperforming prior models on benchmarks for math, coding, science, and visual understanding. O4-mini offers a cheaper, faster alternative that OpenAI says still competes strongly on performance.
For the first time, OpenAI’s reasoning models can invoke tools — web browsing, Python execution, image processing, and image generation — during the chain-of-thought phase. Users can upload blurry photos or whiteboard sketches, and the models will zoom, rotate, or analyze them before responding. The models also execute code inside ChatGPT’s Canvas and can search the web for current events.
The launch comes amid intense competition: Google, Meta, xAI, Anthropic, and DeepSeek have all released reasoning models that match or exceed OpenAI’s earlier o1 series. CEO Sam Altman had signaled in February that OpenAI might not ship o3 into ChatGPT at all, but competitive pressure apparently reversed that decision. The models are available today to ChatGPT Pro, Plus, and Team subscribers, and via OpenAI’s APIs at $10 per million input tokens for o3. OpenAI plans to release an o3-pro version for Pro subscribers in the coming weeks.
One year later — open only if you can handle spoilers
o3 set a new state-of-the-art on several benchmarks before being superseded by GPT-5 later that year, as Altman had hinted. The 'thinking with images' feature became a template for later multimodal reasoning models from Anthropic and Google.