Introducing OpenAI o3 and o4-mini
OpenAI ● Covered by 2 sources
OpenAI just dropped o3 and o4-mini, its newest reasoning models. Line two: these ones can actually use tools like search and code on their own while thinking, not just chat.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI is done treating reasoning and tool use as separate problems. With o3 and o4-mini, the company is fusing its smartest chain-of-thought models with full agentic access to things like web search, Python execution, file analysis, and image generation. That's the real headline here, more than any benchmark bump.
Previous reasoning models like o1 were smart but boxed in. They could think hard about a problem, but they couldn't go fetch a fact, run a script to check their own math, or generate a diagram mid-thought. o3 and o4-mini change that. The models can now decide, on their own, when a tool would help answer a question better, then weave the result back into their reasoning before answering. OpenAI frames this as the models' most capable release yet, and the emphasis on 'full tool access' suggests this is less about raw IQ gains and more about closing the gap between thinking and doing.
o4-mini is the interesting one for anyone watching costs. It's positioned as the smaller, cheaper sibling, but with the same tool-using architecture as o3. That's the pattern OpenAI has leaned on since the o1-mini days: ship a flagship model, then a lighter version that keeps most of the smarts at a fraction of the price. If o4-mini holds up under real usage, it could end up being the workhorse model for developers building agents that need to reason and act without racking up a huge bill.
What's missing from OpenAI's own framing is any detail on how these tool calls are verified or sandboxed, which matters a lot once a model is running code or browsing the web autonomously. Letting a reasoning model decide for itself when to pull a lever is a meaningfully different risk profile than a chatbot that just talks. OpenAI is betting that better reasoning also means better judgment about when to reach for a tool. Whether that bet pays off will show up in how these models behave once they're in the hands of people trying to break them.
My take — AI-written commentary, not fact-checked reporting
I run TLDRocket because I'm tired of AI coverage that's either breathless hype or reflexive doom, so here's the plain read: bolting tool use onto reasoning models is the correct move, and everyone else will copy it within a quarter. My bigger worry is that OpenAI keeps shipping capability faster than it explains the guardrails around autonomous tool calls, and 'trust our judgment' isn't a safety framework, it's a marketing line.
Read more about this at: OpenAI