OpenAI · 2025 · 04 · 16 · Model · ~1 min read
OpenAI launched o3 + o4-mini
What's actually new
- o3 in production. The December 2024 o3 demo wasn't a real product — it was an eval-only preview. April 2025 was the actual release.
- o4-mini. Faster, cheaper, almost as capable as o3 for many tasks.
- Tool use during reasoning. The models can call tools (Python, web search, file access) while they're thinking — not just before or after.
If you want more
Worth knowing
- 'o3' is confusing. Different from o3-preview (Dec 2024), o3-pro (later 2025), and the o3 used inside Deep Research (Feb 2025). Same family, different products.
- Per-call cost climbed meaningfully versus o1. The thinking-time tax is real.
Who should care
Developers who'd been waiting for a real o3 product since December 2024. Researchers needing strong reasoning combined with tool use. Anyone deciding whether to subscribe to ChatGPT Pro versus Plus.
What to do about it
Try o4-mini first — it covers most reasoning tasks at a much lower price than o3. Use o3 for the genuinely hard things you'd be willing to wait two minutes for.
Honest take
o3's actual release closed the loop on a strange four-month gap during which 'o3' was famous but not buyable. The interesting design choice was tool-use-during-thinking — earlier reasoning models could call tools, but only in well-defined steps. o3 weaves them in mid-thought, which sounds small but really changes what kinds of problems it can solve. The whole o-series taught the industry that scaling thinking time keeps paying off where scaling pre-training stopped.
Sources
- OpenAI — Introducing o3 and o4-minivendor
- Simon Willison — o3 + o4-mini first impressionsthird party
- Ethan Mollick — o3 in real workflowsthird party
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com