OpenAI Launches o3 and o4-mini, Its Smartest Models Yet

OpenAI has introduced o3 and o4-mini, which the company describes as "our smartest and most capable models to date with full tool access."

The announcement names two distinct models. o3 sits at the top of OpenAI's capability range. o4-mini is a smaller, presumably faster and cheaper variant aimed at workloads where efficiency matters. The company confirmed neither pricing nor benchmark figures in the announcement itself.

The phrase "full tool access" is the operative detail. It signals that these models can use external tools — web search, code execution, image analysis and similar capabilities — as an integrated part of how they reason and answer, rather than as bolted-on features. That matters because the industry's center of gravity has moved from raw chatbot responses toward agentic systems: models that plan multi-step tasks, call tools along the way, and verify their own outputs.

For OpenAI, the release continues the o-series lineage of reasoning models, which the company has positioned as a separate track from its GPT chat models. Reasoning models spend more compute at inference time to work through problems before answering, a trade-off that has produced strong results on math, coding and science benchmarks in earlier generations.

The stakes are competitive. Google, Anthropic and xAI have all shipped frontier models in recent months, and buyers increasingly judge models on what they can do — browse, code, operate software — not just what they can say. Tool access is now table stakes for that contest, and OpenAI is stating plainly that o3 and o4-mini meet it.

The "mini" tier also reflects an industry-wide pattern: labs pair each flagship with a distilled, lower-cost model so developers can route easy queries cheaply and reserve the flagship for hard problems. o4-mini serves that role here.

OpenAI did not disclose availability windows, API pricing or usage limits in the announcement. Expect those details, along with benchmark comparisons, to follow as developers get hands-on access — and expect the agentic tooling story, not the model names, to decide whether o3 and o4-mini hold their edge.