Introducing o3 and o4-mini

The o3/o4-mini announcement (April 2025) marked OpenAI's acceleration of the reasoning model roadmap — releasing two models that substantially advanced the state of the art on math, coding, and agentic task completion within six months of o1's launch. o3 scored 88% on ARC-AGI-1 (compared to ~85% for humans), effectively closing the gap on a benchmark that had been used to argue current AI systems were far from human-level general reasoning. o4-mini achieved similar performance to o1 at significantly lower cost, making chain-of-thought reasoning economically practical for high-volume applications. Altman's framing for the release emphasized the agentic capabilities: both models were specifically tuned for multi-step task completion where the model plans, executes, and revises over extended sessions rather than responding to single prompts. The announcement also introduced tool use as a first-class capability in the o-series — models could run Python, browse the web, and execute code as part of their reasoning process. This made o3/o4-mini the first OpenAI models genuinely suited for deployment as autonomous agents rather than assistants. The Chollet benchmark result was the most widely discussed aspect externally, but internally Altman cited the agentic tool use as the capability jump that most changed how he thought about AGI timelines.