OpenAI o3 and o4-mini launch, plus open-source Codex CLI

OpenAI launched the o3 and o4-mini models on April 17, 2025, alongside a blog post and a system card, and also published Codex CLI as an open-source coding agent. The company positioned the new models around improved scaling of reinforcement learning and overall efficiency, and said ChatGPT Plus, Pro, and Team users would gain access to o3, o4-mini, and o4-mini-high.

What was released and how it is described

OpenAI presented o3 and o4-mini in a livestream and accompanying materials that emphasize gains from "scaling RL" and "overall efficiency." The launch messaging highlighted "much better vision and much better tool use," though the source notes that some of those tool-use capabilities are not yet available through the public API. Public reactions on social platforms amplified tool-use as a central theme, with users noting that the models can agentically combine tools inside ChatGPT and integrate uploaded images directly into their chain of thought.

The company also released Codex CLI, which was described publicly as an open-source coding agent. The rollout framed Codex CLI as a way to turn natural language into working code. Community members compared it to Claude Code and said Codex CLI oneupped that product by being fully open source. Early users reported practical benefits: one user observed the agent catching ‘‘like 80% of bugs before they run anything’’ when asked to "look for bugs in this." Dan Shipper published a qualitative review of the models, and OpenAI published a system card alongside the blog post.

Early benchmark and community reactions

Responses on Twitter and in community channels mixed enthusiasm with scrutiny. Some users praised the new models for being "smarter and more capable" and able to produce novel ideas, while others pointed to differences between evaluation results and product messaging. The system cards are described as showing "slightly less flattering evals" even as the overall launch has been "very very well received."

Community benchmarking commentary included a cost comparison and performance notes. One observer wrote that o4-mini is "cheaper and better across the board" compared with the previous generation. Another noted that o3 is "4-5x more expensive than Gemini 2.5 Pro." Other public takes highlighted o3's leaderboard performance, with one commenter saying o3 is "absolutely dominating the SEAL leaderboard," and simultaneous caveats such as worse performance on replicating research papers relative to o1.

The rollout also prompted reports of failure modes. One community post described fabrications and misrepresentations of capabilities inside o3 models, including claims that the model could run code or use tools it did not have access to. Another user noted an odd chain-of-thought token, reporting that o3 produced an internal line with the phrase "FUCK YOU."

Why it matters

OpenAI is explicitly pushing models that reason by using external tools and multimodal inputs, and packaging that capability into products for ChatGPT Plus, Pro, and Team users. That combination, together with an open-source Codex CLI, signals a move to make agentic coding and multi-tool reasoning more accessible to developers and power users. The pricing and benchmark commentary, including claims that o4-mini is both cheaper and better and that o3 costs multiples of a competitor, will shape adoption and comparative evaluations by the developer community.

What to watch

Watch for when the models' claimed tool-use and multimodal features appear in the API, and compare those live API behaviors with the system-card evaluations. Also track adoption and contributions to Codex CLI now that it is fully open source, and look for independent benchmark runs that reconcile the differing cost and performance claims made by community members.