OpenAI releases o3 and o4-mini

Background

The name o3 first became public on 20 December 2024, when OpenAI announced the model. It skipped the number that would have followed naturally — o2 — because "O2" was already a trademark held by another company, so the release moved straight to o3^[1]. Nearly four months passed between that announcement and the point where the models were actually available to users.

What was released

On 16 April 2025 OpenAI released o3 and o4-mini^[1]. OpenAI billed the pair as its smartest models to date, and said that for the first time its reasoning models could agentically use and combine every tool inside ChatGPT^[3]. Reasoning models are ones that work through a longer internal chain of thought before producing an answer; the new claim is that such a model can also decide on its own to call a search, a code interpreter or another built-in tool, and chain several of them together while working on a task. The "mini" name marks the lighter member of the family.

Benchmarks

OpenAI reported o3 at 71.7% on SWE-bench Verified, a benchmark built from real software issues in open repositories, where a model has to produce a patch that makes failing tests pass. It also reported a Codeforces Elo of 2727 for the model; Codeforces is a competitive-programming site, and the Elo figure comes from the same rating system used for human contestants. Both numbers point at code and at problems that require many steps of reasoning before an answer appears^[1].

Codex CLI

The same day brought a second, smaller release: Codex CLI, an open-source coding agent that runs in a terminal^[2]. It is used from the command line, where a developer can point it at a local repository and let it read, edit and run code. OpenAI published it as fully open source at github.com/openai/codex, and paired the release with a $1 million initiative that awards grants in $25,000 increments of API credits^[3]. The money is aimed at people building on top of the tool rather than at OpenAI's own use of it.

Safety

Alongside the launch OpenAI published a system card describing how the two models were evaluated. Under the Preparedness Framework, the company reported that both models stayed below the "High" threshold in three categories: biological and chemical risk, cybersecurity, and AI self-improvement^[3]. Staying below a threshold is a statement about measured evaluations, not a claim that no risk exists.

Who could use them

Availability changed on launch day. For Plus, Pro and Team subscribers, o3, o4-mini and o4-mini-high took the place of o1, o3-mini and o3-mini-high in the model selector; Enterprise and Edu accounts followed a week later^[3]. The substitution is worth noting: the previous generation did not sit alongside the new one as a separate option in those menus, it was replaced.

What came after

A stronger variant, o3-pro, followed on 10 June 2025^[1], extending the same line of models.