OpenAI releases o3 and o4-mini, and also open sources two projects

Written by

Caleb Hayes

Updated on:June-30th-2025

Last night OpenAI released two new artificial intelligence models: o3 and o4-mini, which are the first to give AI the ability to "think by looking at pictures".

Model comparison

o3: The most powerful reasoning model, good at complex tasks.

o4-mini: Lightweight, fast, cost-effective, and suitable for daily tasks.

Pricing strategy: o3 is 33% cheaper than its predecessor o1, while o4-mini maintains the original price, reflecting OpenAI's attempt to balance performance and cost.

User selection suggestions

It is recommended to select o3 for scientific research/development: advanced reasoning, tool calls, or processing of complex images (such as scientific research charts

Choose o4-mini for daily/light tasks: cost-sensitive scenarios that require basic visual capabilities (such as document parsing)

Use with caution: Fact-checking tasks require manual verification due to the increased rate of hallucinations.

Core breakthrough

For the first time, image analysis is integrated into the thinking chain (such as understanding diagrams in scientific papers and solving geometric figure problems), which solves the pain point of traditional multimodal models that can only "speak with pictures" but cannot make deep reasoning.

Functional upgrade: o3 supports Python, browser, API call and other tool chains, and the task completion rate is significantly improved (for example, after connecting to the tool, the AIME accuracy rate has increased from 91.6% to 95.2%)

Immediate experience: ChatGPT Plus/Pro/Team users can experience o3 and o4-mini, and ordinary users can try o4-mini for free.

Open Source Codex Agent Project

OpenAI Live also open-sourced a Codex CLI local coding agent project that can directly convert natural language into executable code.

Just tell Codex CLI what you want to build, fix, or interpret, and it will bring your ideas to life.

Codex CLI is compatible with all OpenAI models, including o3, o4‑mini, and GPT‑4.1.

Github: openai/codex

openai-fm website open source

This morning I also saw that OpenAI also open-sourced the openai.fm website, the technology stack is Nextjs + TailwindCSS + ShadcnUI

GitHub: openai/openai-fm

Origin of Codex

In 2020, OpenAI released GPT-3, which is impressive in its powerful language generation capabilities, but its interactivity and logical coherence are not as good as the later ChatGPT-3.5.

In order to explore the potential of LLM in programming, OpenAI developed an LLM for assisted programming based on the GPT-3 architecture and trained it with a large amount of code data, and named it Codex.

In 2021, GitHub released Copilot, which was supported by Codex LLM. During the training process of Codex, researchers found that through a large amount of code learning, LLM's logical ability was significantly enhanced.

So, OpenAI adjusted the training method, combined massive text, code and RLHF, and finally developed ChatGPT-3.5, which amazed the world at the end of 2022.

Codex is the beginning of this series of innovations, and the open source gadget released today is named Codex, as a tribute to this pioneering model.