THE BIGGEST name in artificial intelligence finds itself in an unfamiliar position: chasing. OpenAI, the company that made "ChatGPT" synonymous with AI itself, has spent much of the past year scrambling to close a gap in what may be the most lucrative product category the industry has yet produced. Its rival Anthropic, founded by OpenAI defectors, commands roughly $2.5 billion in annualized revenue from Claude Code alone. OpenAI's competing product, Codex, was generating just over $1 billion as of late January — a respectable sum, albeit one that makes the self-proclaimed leader of the AI revolution look like a follower.
The backstory is one of strategic neglect. OpenAI built the original Codex in 2021 and promptly handed the commercial opportunity to Microsoft, which used it to power GitHub Copilot. When ChatGPT exploded in late 2022, every internal project bent toward the consumer chatbot. For years, OpenAI had no dedicated coding product team. The prevailing logic was that GitHub Copilot had the sector "covered," and that multimodal models — AI that could see, hear, and manipulate a screen — represented the true path to artificial general intelligence. Anthropic, meanwhile, was quietly training its models on messy real-world code repositories, a decision OpenAI's own president, Greg Brockman, has publicly acknowledged as a lesson his company was "delayed on."
Terminal velocity
Yet the gap may be narrowing faster than the headline numbers suggest. By September 2025, Codex was attracting just 5 percent of Claude Code's usage. By January 2026, that figure had climbed to roughly 40 percent, fueled by the launch of GPT-5.2 and an aggressive sprint team that OpenAI assembled in March 2025. The company threw a Super Bowl ad behind Codex (not ChatGPT), signed enterprise customers including Cisco, and leaned on its greatest distribution asset: the fact that hundreds of millions of people already use ChatGPT. "Companies want to use technologies their workers are already familiar with," says Fidji Simo, OpenAI's CEO of applications.
But distribution advantages have a shelf life in markets where the product itself is evolving weekly. OpenAI's attempt to buy the coding startup Windsurf for $3 billion — a deal that would have delivered an established product, a seasoned team, and enterprise customers in one stroke — collapsed in July after Microsoft, protective of GitHub Copilot's position, demanded access to Windsurf's intellectual property. Google hired Windsurf's founders; the rest went to Cognition. The episode illustrates a structural tension that OpenAI's competitors do not share: every ambitious product move must navigate a partnership with Microsoft that is simultaneously OpenAI's greatest asset and its tightest constraint.
The economics of this race, meanwhile, are punishing for both sides. Developers report that their $200-per-month plans routinely deliver over $1,000 in usage — a subsidy designed to embed AI coding agents into workplace habits before switching to usage-based pricing. It is a land grab dressed up as a product launch, and neither company appears willing to blink first. The stakes justify the spend: Altman reckons AI coding is "one of these rare multitrillion-dollar markets," and describes Codex as "probably the most likely path" to AGI. Whether that framing reflects genuine conviction or fundraising gravity is left as an exercise for the reader.
The societal tremors are already registering. The Wall Street Journal attributed a $1 trillion tech stock sell-off last month to fears that Claude Code could render swathes of software engineering obsolete. IBM suffered its worst trading day in 25 years after Anthropic announced Claude Code could modernize legacy COBOL systems. Cisco's chief product officer has told employees they will lose their jobs not for using AI coding tools, but for refusing to. The discomfort extends even to OpenAI's inner circle. Brockman, an engineer's engineer, concedes that commanding a "fleet of hundreds of thousands of agents" can make you "feel like you're losing your pulse on the problem."
Safety concerns hover at the edges. The Midas Project accused OpenAI of weakening its safety commitments with GPT-5.3-Codex; OpenAI disputes the characterization. Altman himself declined to set up a viral coding agent for a friend, deeming it "clearly not a good idea yet." Weeks later, OpenAI hired its creator.
The coding agent war is entering a phase where neither raw model capability nor brand recognition alone will decide the winner. What matters now is execution velocity, enterprise trust, and the willingness to lose money in precisely the right places. OpenAI has closed ground with remarkable speed; whether it can overtake Anthropic depends on whether catching up and leading require the same skills. History suggests they do not. ■
For more, join 75,000 subscribers getting tech's favorite brief here
