Inside OpenAI’s fast-growing Codex: The people building the AI that codes alongside you

OpenAI’s Codex AI coding assistant is having a growth spurt. OpenAI tells Fast Company that its weekly active users have tripled since the start of the year, while overall usage (measured in tokens) has increased fivefold. The surge is likely driven by the release of new models—GPT-5.2 last December and GPT-5.3-Codex in early February—as well as the launch of Codex’s app version a few weeks ago. OpenAI says the app has been downloaded more than a million times. Across all access points—including the cloud, app, and command line—more than a million developers and other users now rely on Codex at least once a week, according to the company.

Generating computer code has emerged as one of the first AI applications making a measurable impact in business. But tools like Codex and Anthropic’s Claude Code have evolved far beyond simple code generators. Powered by more capable models, they function more like assistant engineers—able to converse with developers in plain language about a new software project and iteratively develop a plan. The agent can then execute that plan, which may include analyzing a broader codebase, writing and revising code, conducting research, running tests, and producing documentation. When finished, it can explain its reasoning and the decisions it made to the human engineer.

More importantly, Codex has evolved into an agentic platform, where multiple agents can carry out many of these tasks simultaneously across different pieces of a software project. They can hunt for bugs, for example, while an engineer reviews progress, focuses on another assignment, or steps away for lunch. Peter Steinberger, the OpenClaw creator and an elite-level coder, calls this new mode of working “agentic engineering.”

The tools have evolved quickly. Codex and Claude Code both launched in the first half of 2025. OpenAI had previously introduced a Codex model in 2021—the system that powered the early AI coding assistant GitHub Copilot—but the Codex coding assistant that exists today debuted in May 2025.

Thibault Sottiaux, who leads the Codex group at OpenAI, says the product got a major boost with the December 2025 release of the GPT-5.2 model, which he says can hold more project data in memory and reason over it more effectively than earlier versions. “The model was more reliable—working by itself autonomously and reaching really good results,” he tells Fast Company.

Codex’s user base broadened again with the February 2 release of the Codex desktop app for Mac, which OpenAI describes as a “command center” where users can deploy and manage multiple agents. The company says more than half a million people are now accessing Codex through ChatGPT’s Free and Go subscription tiers, and it believes many of them are non-coders, since power users typically rely on higher-priced plans that offer greater usage limits and faster speeds.

The biggest bang came with the February 5th launch of GPT‑5.3‑Codex, which substantially improved Codex’s coding chops, as well as its capacity for reasoning its way through complex, long-running tasks that involve research and tool use. In X posts and Reddit discussions many developers raved about the tool’s capacity for quickly writing usable code for real-world projects, often on the first try. 

Codex vs. Claude Code

Many of the AI coding agents on the market are powered by third-party models, but OpenAI and Anthropic, along with Google and its Gemini Code Assist product, are each trying to leverage the strengths of their own frontier large language models to deliver the most capable and reliable coding tool. OpenAI’s Codex and Anthropic’s Claude Code share some broad similarities. Both can build large features or even entire apps based on plain-English conversations with a user. Both also allow developers to break complex projects into subtasks and assign those to agents.

But there are differences. One major distinction is the look and feel, or what some describe as the “personality,” of the tools. Steinberger says Claude Code is more conversational and iterative than Codex. It includes, for example, a dedicated planning phase before any code is written. Codex, by contrast, does not formally separate planning and coding and instead tends to dive directly into the codebase to gather context and begin working. Steinberger (comically) described the difference this way on a recent episode of Lex Fridman’s podcast: “Opus [Anthropic’s flagship Claude model] is like the coworker that is a little silly sometimes, but it’s really funny and you keep him around,” he said, “and Codex is like the weirdo in the corner that you don’t wanna talk to, but is reliable and gets shit done.” (OpenAI has since acquired Steinberger’s OpenClaw agent platform, and Steinberger now works at OpenAI.)  

“The pragmatic personality has always been the personality that we have on Codex,” Sottiaux says, “which is very much focused on having the model point out flaws and being as correct as possible when it comes to discussing something and being a very reliable tool.”

The personality and interaction habits of AI agents can reflect the markets they’re designed to serve. “We were just really focused on this professional software engineering audience and . . . on getting to a powerful agent that can do tasks independently,” Codex product manager Alex Embiricos says.

But those target markets can shift. Embiricos says that while a pragmatic approach works well for experienced developers, less experienced or first-time coders may prefer a more empathetic, conversational interface. And that audience is growing as Codex evolves into a tool for general information work. That’s one reason the Codex team decided to give users more choice within the app.

“In January we said ‘Okay, we’re doing great on intelligence; obviously there’s more to do, but now we’re going to actually spend a few more cycles on personality,'” Embiricos says. With the arrival of the GPT-5.3-Codex model, Codex now offers the default “pragmatic” personality as well as  a new “empathetic” or “friendly” mode, which is designed to be more conversational and interactive.

Why are AI models so good at coding?

At the most basic level, computer code is made up of words, the same kind of data large language models are designed to process. And because the people building AI models are themselves programmers, they have strong incentives to make their systems excel at coding.

Computer code is also in training and evaluating models. While there’s creativity involved in software engineering, code ultimately either works or it doesn’t. That creates a large supply of training examples with clear right and wrong answers. “There’s lots and lots of examples out there with a problem statement and a solution, and being able to tell whether the solution is correct or not,” Sottiaux explains. “So you can at the very least use that for evaluations to understand the performance of models over time, and drive that performance up.”

Codex is still a young product, and OpenAI says it’s improving quickly. But it’s still a work in progress, and in the weeks since the GPT-3.5-Codex model upgrade, developers have reported problems in some coding scenarios. Some users say GPT-5.3-Codex can lose focus during long or complex tasks, get stuck in loops, freeze, or repeatedly ask for approval instead of completing work. Others say it can hallucinate plausible-looking code, especially in front-end fixes, that doesn’t actually work. These accounts are anecdotal and not systematically measured, but they underscore a common practice among developers of keeping AI-generated code separate from production systems until it’s reviewed.

The Codex team has been focused on identifying and removing near-term bottlenecks that limit usefulness, according to research scientist Amelia Glaese, who leads development of the models underneath Codex. “You know, three months ago, people were using Codex, but they were using it a lot less than they are using it now,” Glaese adds. “There were changes that we made two months ago and two weeks ago that made it so much more useful to people.”

At the same time, tools like Codex and Claude Code require developers to adapt. Working with an AI coding assistant is a different mode of software engineering, one that involves guiding and collaborating with an agent rather than writing every line directly. “It’s not the case that there’s like one right way of solving an engineering problem,” Sottiaux says. “It’s all a question of trade-offs and exploring those trade-offs, and so when you have an agent that’s capable of helping you explore those trade-offs, it’s a very useful tool for an engineer.”

Increasingly, these assistants are capable of contributing to the development of the next generation of AI models themselves. If AI systems eventually handle more of the process of building, training, evaluating, and deploying models, the pace of performance improvements could accelerate significantly.

Not just coding

Both Codex and Claude Code are evolving into tools for general information work. Anthropic has drawn significant attention as it rolls out new Claude Cowork plugins (bundles of information-work skills) such as for sales, finance, and legal work. Cowork appears as a separate tab, alongside Claude Code, within the Claude chatbot interface. Anthropic’s skills announcement helped trigger a sell-off in software stocks, reflecting investor fears that traditional software-as-a-service products could be displaced by AI tools sooner than expected.

OpenAI is also adding information-work skills to Codex, if more quietly. “Skills bundle instructions, resources, and scripts so Codex can reliably connect to tools, run workflows, and complete tasks according to your team’s preferences,” the company wrote in the blog post announcing the GPT-5.3-Codex model. The Codex app includes a dedicated interface for creating and managing these skills. OpenAI already has a large and expanding portfolio of products, but it considers Codex important enough to feature in its “You Can Just Build Things” Super Bowl ad this year.

Glaese, for her part, points out that software engineers themselves have a natural incentive to expand Codex beyond coding tasks. Much of their workday involves general information work rather than writing code. “We have to do research, we have to understand the market, we have to read news, we have team meetings, we do performance reviews—we do all of the things that people who don’t code also do,” she says.

The glaring question around agents like Codex and Claude Code is how they will affect human jobs, especially those of younger engineers. OpenAI wants its agent to behave like a talented assistant engineer but stops short of saying it will replace people. Instead, Sottiaux sees coding agents as a way to expand how teams approach problems and develop new ideas, particularly when less experienced engineers use them to experiment and push beyond conventional approaches. “And then they come up with completely new ideas that you might not have if you anchor too much on your decades of experience,” he says.

View the full article