OpenAI says its seeing breakout growth for its AI coding tool Codex, even as controversy over the company’s agreement to supply AI to the Pentagon has derailed the public messaging around Codex’s momentum and resulted in some consumers boycotting its ChatGPT product.
Since early February, when OpenAI launched GPT-5.3 Codex, the latest and most capable version of its coding agent, more than 1 million people have downloaded the codex desktop app and Codex now boasts more than 1 million active weekly users, a figure that has tripled with the release of the new model, according to the company. It also said that usage, as measured in the number of tokens, or portions of text, that Codex is processing per week has grown by a factor of five. Companies including Cisco, Nvidia, Ramp, Rakuten, and Harvey have rolled Codex out across their developer teams, according to OpenAI.
In an interview in London last week, before the controversy over the Pentagon deal erupted, Thibault Sottiaux, the head of OpenAI’s Codex product, laid out the company’s ambitions to use Codex as a mechanism to bring agents to the enterprise in domains beyond coding.
Code as a tool that uses other tools
He described Codex as “becoming the standard agent” that OpenAI plans to expand across enterprise deployments, including for non-technical workers—though he acknowledged there is still significant work to do on security, managed deployments, and on-premises offerings.
“Fundamentally, the agent is composed of the model, and then the harness that enables us to access your file system, make changes,” Sottiaux said. “There’s very little that is specific to coding.” A harness is a set of systems around an AI model that defines and controls how it can use tools, how it remembers things, and what guardrails it has.
Sottiaux described Codex’s core training as focused on “instruction following, understanding large amounts of data, finding its own context, and navigating the world in order to make decisions on its actions”—capabilities, he argued, that are as useful outside of code as within it.
The key insight, according to Sottiaux, is that code can be used to automate the use of other software, such as crunching data in spreadsheets or building a financial model from data found in disparate documents. “If we manage to sandbox it properly and make it safe for non-technical users, then suddenly you can bring the power of coding agents to billions of users,” Sottiaux said. Codex already employs “Skills”—shareable, composable text-based instruction sets that steer agent behavior—and Sottiaux said marketplaces for these Skills are beginning to emerge.
The strategy Thiabult outlined for Codex is similar to the one pursued by OpenAI’s rival Anthropic, which has seen viral growth of its Claude Code product among software developers but which has also sought to position Claude Code as a tool that other professionals can use to spin up AI agents too. It released Claude Cowork, a separate product, designed specifically to help people use AI agents to control common work software products, such as spreadsheets, email, and calendar apps.
The rush to roll out universal AI agents
Numerous companies are rushing to move into the space for AI agents, particularly following the viral popularity of OpenClaw, an open-source AI agent harness that can be used with any AI model as the underlying “brain.” Following the explosive usage of OpenClaw, Perplexity in late February debuted an agentic AI system called Computer, which is a cloud-based system that orchestrates 19 different AI models to execute complex workflows. Microsoft has also launched Copilot Tasks, which is a similar AI agent and harness product. Meanwhile, OpenAI hired Peter Steinberger, the independent developer who built OpenClaw, although OpenClaw itself will continue to be run as an open source project under the auspices of a foundation Steinberger set up.
Sottiaux said he was excited to work with Steinberger going forward. He called OpenClaw “a magical experience” and “a glimpse of the future,” but added that “it’s not something that everyone should just run on their machine unchecked.” Security researchers had found a number of serious vulnerabilities in using OpenClaw, and several users reported that the system had been subject to “prompt injection” attacks (where someone feeds an AI agent malicious instructions) that resulted in data breaches. Other users reported that OpenClaw could undertake unintended and damaging actions, such as deleting email accounts and other data.
OpenAI wants to take elements of OpenClaw’s approach but “package it in a way where everyone would be able to benefit from an always-on personal agent” and, hopefully, have much better security and safeguards, Sottiaux said.
Asked whether OpenAI’s widely-reported internal “Code Red” had changed how his Codex team operated, Sottiaux was dismissive. “[Codex] is a really small team, and we’re firing on all cylinders,” he said. “We’ve been just really saying no to a lot of things from the get-go, and working on things that we think we’re uniquely good at doing, and then shipping continuously.”
Pentagon controversy eclipses Codex’s breakout
The story around Codex’s surge has been largely eclipsed by the drama surrounding OpenAI’s controversial decision to make a deal with the Pentagon that will allow the Department of War to use OpenAI’s models in classified networks.
The deal, announced on Feb. 28, followed a breakdown of negotiations over a similar contract between the Department of War and Anthropic and hours after Secretary of War Pete Hegseth designated Anthropic a “supply-chain risk” in retaliation for Anthropic’s insisting that it would not sign a contract without specific limitations on the military using its Claude models for mass surveillance of Americans or to control autonomous weapons.
OpenAI CEO Sam Altman had previously said he supported Anthropic’s red lines and he said that the company’s own contract with the Department of War included language designed to create the same limits, but legal experts questioned how effective that language would be. Altman later acknowledged the contract was “opportunistic and sloppy,” and the company has since renegotiated parts of the agreement.
In the meantime, however, OpenAI has faced criticism from many quarters, including from among its own employees. Some have called for a consumer boycott of OpenAI’s ChatGPT product and Anthropic’s Claude has surged past ChatGPT to become the No. 1 free app on Apple’s App Store, propelled by an online campaign urging users to switch.
Whether this consumer revolt is translating into any meaningful attrition among developers using Codex is not clear, however. App store rankings reflect downloads of consumer chatbot apps—a different market from the professional developer audience that drives Codex usage.
News of OpenAI’s Codex’s growth also comes amid reports of surging business adoption for Anthropic’s products. Data released by Ramp, a software company that handles expense management, show that Anthropic’s marketshare of business AI chatbot invoices has climbed to more than 60% in February, from just over 10% a year earlier. Meanwhile, Ramp’s figures showed OpenAI’s business marketshare falling to about 35%, down from almost 90% the year before. Anthropic CEO Dario Amodei also told a conference this week that his company was operating at a $19 billion annualized revenue run rate, a figure that climbed by $6 billion in February.
It is unclear if OpenAI’s reported momentum for Codex can help arrest any decline in other areas of its business or reverse the narrative that it is losing enterprise marketshare to Anthropic.
