OpenAI says its AI coding tool Codex is experiencing exponential growth, even as controversy over its agreement to supply AI to the Department of Defense derailed public messaging around Codex’s momentum and resulted in some consumers boycotting the company’s ChatGPT product.
Since early February, when OpenAI announced 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 weekly active users, a number the company says has tripled with the release of the new model. It also said that usage, measured by the number of tokens or pieces of text that Codex processes per week, has increased five times. According to OpenAI, companies like Cisco, Nvidia, Ramp, Rakuten, and Harvey are deploying Codex across their developer teams.
In an interview in London last week, before the controversy over the Pentagon deal erupted, Thibault Sautiaux, OpenAI’s head of Codex products, explained the company’s ambitions to use Codex as a mechanism to bring agents into enterprises in areas beyond coding.
Codex as a tool that uses other tools
He explained that Codex, which OpenAI plans to extend across enterprise deployments, including non-technical employees, is “becoming the standard agent,” but acknowledged that significant work remains regarding security, managed deployment, and on-premises products.
“Essentially, an agent consists of a model and a harness that allows it to access the file system and make changes,” Sottiaux says. “There’s very little that’s specific to coding.” A harness is a set of systems around an AI model that defines and controls how tools are used, how things are remembered, and what the guardrails are.
Sotiaux explained that Codex’s core training focuses on “following orders, understanding large amounts of data, finding unique context, and navigating the world to determine behavior.” These abilities, he argued, are useful not only within the code but also externally.
A key insight, Sottiaux said, is that the code can be used to automate the use of other software, such as processing data in spreadsheets or building financial models from data contained in different documents. “If you sandbox it properly and make it safe for non-technical users, all of a sudden you can bring the power of a coding agent to billions of users,” Sottiaux said. Codex already employs “skills” – shareable, configurable, text-based sets of instructions that control agent behavior – and Sotiaux said a market for these skills is starting to emerge.
The strategy Sottiaux outlined for Codex is similar to that pursued by OpenAI rival Anthropic. While its Claude Code product is rapidly gaining popularity among software developers, Anthropic is positioning Claude Code as a tool that other professionals can use to launch AI agents. The company released Claude Cowork, another product specifically designed to help people use AI agents to control common business software products like spreadsheets, email, and calendar apps.
Accelerate deployment of universal AI agents
Many companies are rushing to enter the AI agent space, especially with the rapid popularity of OpenClaw, an open-source AI agent harness that can be used as the underlying “brain” in any AI model. Following the explosive use of OpenClaw, Perplexity debuted an agent AI system called Computer in late February. It is a cloud-based system that coordinates 19 different AI models to perform complex workflows. Microsoft also launched Copilot Tasks, a similar AI agent and harness product. Meanwhile, OpenAI hired Peter Steinberger, the independent developer who built OpenClaw, which itself will continue to operate as an open source project under the auspices of the foundation founded by Steinberger.
Sotiaux said he is excited to be working with Steinberger in the future. He called OpenClaw a “magical experience” and “a glimpse into the future,” but added, “It’s not something anyone should be running on their machine unchecked.” Security researchers have discovered a number of serious vulnerabilities in the use of OpenClaw, and several users have reported that their systems have been subject to “prompt injection” attacks (where someone gives malicious instructions to an AI agent), resulting in data breaches. Other users reported that OpenClaw could take 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 that everyone can benefit from an always-on personal agent,” and hopefully with much better security and protection measures, Sotiaux said.
When asked if OpenAI’s widely reported internal “Code Red” has changed the way his Codex team operates, Sotiaux was dismissive. “[Codex] “We’re a really small team firing on all cylinders. We’ve really said no to a lot of things from the beginning, worked on what we think we’re uniquely good at, and then continued to ship,” he said.
Pentagon controversy upends Codex breakthrough
The story surrounding Codex’s proliferation has been largely overshadowed by the drama surrounding OpenAI’s controversial decision to enter into an agreement with the Department of Defense that would allow the Department of the Army to use OpenAI’s models on classified networks.
The agreement was announced on February 28, hours after negotiations between the Army Department and Anthropic for a similar contract broke down, with Secretary of the Army Pete Hegseth designating Anthropic as a “supply chain risk,” and in retaliation for Anthropic’s insistence that it would not sign the contract unless there were specific restrictions on the U.S. military’s use of the Claude model for mass surveillance and autonomous weapons management.
OpenAI CEO Sam Altman has previously said he supports Anthropic’s redlining, and noted that the company’s contract with the Department of the Army includes language that seeks to create similar restrictions, but legal experts question how effective that language would be. Mr. Altman later acknowledged that the deal was “opportunistic and sloppy,” and the company later renegotiated parts of it.
But during that time, OpenAI has faced criticism from many quarters, including from its own employees. While some have called for a consumer boycott of OpenAI’s ChatGPT product, Anthropic’s Claude has overtaken ChatGPT as the No. 1 free app on Apple’s App Store, driven by an online campaign encouraging users to make the switch.
However, it is unclear whether this consumer revolt has translated into a meaningful decline among developers using Codex. App Store rankings reflect consumer chatbot app downloads. This is a different market than the professional developer audience that drives the use of Codex.
The news about OpenAI Codex’s growth also comes amid reports that business adoption of Anthropic products is surging. Anthropic’s market share in business AI chatbot invoicing rose from just over 10% a year ago to more than 60% in February, according to data released by expense management software company Ramp. Meanwhile, OpenAI’s business market share has fallen to around 35% from almost 90% a year ago, according to Ramp statistics. Anthropic CEO Dario Amodei also said at a conference this week that the company operates on $19 billion in annual revenue, a figure that increased by $6 billion in February.
It’s unclear whether OpenAI’s reported momentum with Codex will help stem the decline in other areas of the company’s business, or reverse the narrative that it’s losing enterprise market share to Anthropic.
