The AI Tool Built to Replace Programmers Is Winning Over Everyone Else

OpenAI built Codex to replace programmers. The programmers did not fully embrace it. Now OpenAI is trying to sell it to everyone else, and the reason matters more than the product launch itself.

Codex reached 5 million weekly active users this week, up more than sixfold since the desktop app launched in February. The more interesting number is the one OpenAI quietly included in the same announcement: knowledge workers now represent roughly 20 percent of users and are growing more than three times as fast as the developer base — though that figure comes from OpenAI's own research and has not been independently audited.

That is the market OpenAI is now chasing deliberately. On Tuesday it released six role-specific plugins for Codex: data analytics, creative production, sales, product design, public equity investing, and investment banking. Each bundles integrations with the apps those workers already use — 62 apps total, 110 skills — configured to work out of the box without code. The company also launched a new enterprise sales entity called the OpenAI Deployment Company three weeks prior, backed by more than $4 billion in funding from global investment firms.

The timing is not accidental. Both OpenAI and Anthropic are preparing to go public, and the race to sign up knowledge workers at scale is now an IPO competitiveness question, not just a product strategy question.

Anthropic arrived at that party first. Among developers, Claude Code is still preferred — Ars Technica put it plainly in March: talk to developers and you find more Claude Code users than Codex users. OpenAI introduced plugin support for Codex only in March, trailing Claude Code's ecosystem by months.

So OpenAI is doing what large software companies do when they cannot out-compete on product: buying distribution. The company brought in Accenture, Capgemini, and PwC as consulting partners to sell Codex to enterprise clients. Whether those partnerships have produced actual contracts or only pipeline is not public. OpenAI will not say how many of the new plugins are live versus preview. The knowledge worker growth figures come from an internal research report the company has not made freely available.

There is a deeper question buried underneath the competitive story. Professional expertise has always developed through grunt work. Junior lawyers do document review until they understand what a contract actually does. Junior analysts build models until they learn what the numbers mean. Radiology residents spend years staring at scans before they develop the pattern recognition that distinguishes a shadow from a tumor. The entry-level work is not incidental to professional training — it is the mechanism through which expertise gets built.

Codex automates exactly that kind of work: drafting reports, running analyses, preparing comp sets, summarizing meetings. If those tasks disappear from the junior professional's desk before they have learned what the senior professional knows, something breaks in the pipeline. This is not a hypothetical concern — legal technology researchers have documented how early document review automation reduced the exposure of junior associates to case files they would once have had to read. Radiology is further along in this process, with some training programs reporting that residents arrive with less direct image interpretation experience than their predecessors. Whether AI tools that assist rather than replace will preserve the learning function, or whether the economics of efficiency will push firms to automate the work entirely, is a question nobody in the industry has answered yet.

The competitive dynamic matters for readers beyond the AI industry. OpenAI is valued at $852 billion — the largest private valuation in history — on the strength of roughly $24 to $25 billion in annualized revenue. Anthropic is not far behind. Both are arguing to public market investors that enterprise AI adoption is still early and still growing. The knowledge worker adoption numbers OpenAI published this week are central to that thesis. If those numbers are real, they strengthen the bull case for both companies. If they are concentrated in a few large pilot deployments that have not yet converted to revenue, the picture is different.

The honest version of the story is this: OpenAI built a tool that programmers did not fully adopt and non-programmers did. Now it is racing to see whether enterprise sales and plugin ecosystems can turn that unexpected adoption into the growth narrative it needs for an IPO. What it has not yet grappled with is what happens to the people who would have learned to be the next generation of bankers, analysts, and lawyers if the work that built them gets automated away first.