o3, o4-mini Are Better Than Greg Brockman at Navigating OpenAI’s Codebases

OpenAI has released two new models, o3 and o4-mini, along with a new code generation tool called Codex CLI, marking the latest advancement in its o-series. These models represent a shift in ChatGPT’s capabilities by combining deeper reasoning with full tool access. 

The company calls them “the smartest models” they have released to date, which are able to use all tools within ChatGPT—including web browsing, Python, file analysis, and image generation—in a single workflow.

“These are the first models where top scientists tell us they produce legitimately good and useful novel ideas,” said OpenAI president Greg Brockman during the live stream. The models have shown “great results in law’ and the ability to generate innovative system architecture ideas.

“We’ve trained them to use tools, which is not something that we had done with our previous reasoning models. They actually use these tools in their chain of thought as they’re trying to solve a hard problem,” said Brockman, adding that in one instance, o3 used 600 tool calls in a row trying to solve a really hard task.

“I found that these models are actually better than I am at navigating through our OpenAI codebase, which is really useful,” he added. 

o3 is described as OpenAI’s most powerful reasoning model so far, improving performance across coding, math, science, and visual perception. It sets new benchmarks on Codeforces, SWE-bench, and MMMU without relying on model-specific scaffolds. 

OpenAI said that “o3 makes 20 per cent fewer major errors than OpenAI o1 on difficult, real-world tasks”, with improved performance noted in domains like programming and creative ideation.

o4-mini is a smaller, cost-efficient model designed for high-throughput use. It delivers strong results on tasks requiring math, coding, and visual understanding. It outperforms its predecessor on the AIME 2024 and 2025 competitions and in non-STEM domains such as data science. “Thanks to its efficiency, o4-mini supports significantly higher usage limits than o3,” OpenAI said.

Both models demonstrate improved instruction following and produce more verifiable responses, aided by access to up-to-date web sources. The models also exhibit more natural and conversational interactions, drawing from past exchanges to personalise responses.

Access to o3 and o4-mini is rolling out on Thursday for ChatGPT Plus, Pro, and Team users. Enterprise and education plans will follow in a week. Developers can use both models via the Chat Completions and Responses APIs. An o3-pro version is expected soon.

The new releases are part of OpenAI’s broader strategy to develop agentic AI systems. “These models are trained to reason about when and how to use tools to produce detailed and thoughtful answers in the right output formats, typically in under a minute,” the company explained. 

This enables multi-step task execution, such as generating forecasts based on live data, building visual explanations, or interpreting complex images and diagrams.

OpenAI’s Code Generation Tool

Codex CLI is a lightweight coding agent that brings multimodal reasoning to the command line. Users can combine screenshots or sketches with local code context, and the tool is now open source.

OpenAI researchers Fouad Matin and Michael Bolin introduced Codex CLI, describing it as a “lightweight interface to connect our models to our users and their computers.” The tool allows users to interact with the models directly on their local machines, using features like multimodal reasoning. A live demonstration showcased Codex CLI reimplementing an image-to-ASCII art generator from a screenshot, even incorporating a webcam feed.

A $1 million grant programme has been launched to support projects using Codex CLI and OpenAI models, with proposals accepted in $25,000 increments.

The new models also reflect safety improvements. OpenAI noted that it completely rebuilt its safety training data, particularly for areas like biological threats and jailbreaks. A new reasoning LLM monitor flagged 99% of risky conversations in internal red-teaming tests.