OpenAI's Own Staff Abandoned ChatGPT for Agents. Here's the Data.

The Internal Signal OpenAI Just Made Public

OpenAI published an economic research paper on 25 June 2026 documenting how its own workforce shifted from ChatGPT to Codex, its agentic AI tool. By June 2026, Codex accounted for 99.8% of all output tokens generated inside OpenAI, and every department including Legal, Finance, and Recruiting had adopted it as their primary AI tool.

That is not a product announcement. It is a behavioural signal from the organisation best positioned to observe what capable AI actually does to knowledge work at scale. Growth leaders should read it that way.

What Happened, and When

OpenAI released Codex to the public and tracked adoption from August 2025 onwards. Through that month, the average OpenAI employee spent less than 10% of their AI usage on Codex. ChatGPT remained the default.

The crossover happened in stages. Engineering shifted first, with the average engineer generating the majority of their output tokens via Codex by December 2025. Legal, Finance, and Recruiting followed around April 2026, but their transitions were faster and steeper. By June 2026, the median Research employee was using Codex at 56 times the volume they were in November 2025. Customer Support rose 32x. Engineering rose 27x.

The chart below, drawn directly from OpenAI's paper, shows how each department crossed the 50% threshold at different points but all arrived at the same destination.

Table 1: Codex adoption timeline by department at OpenAI (share of monthly output tokens on Codex, June 2026)

Department	Majority Codex usage crossed	Share of output tokens on Codex (Jun 2026)	Growth vs Nov 2025 (median active user)
Engineering	December 2025	99%	27x
Finance	April 2026	91%	Not separately stated
Recruiting	April 2026	89%	Not separately stated
Legal	April 2026	88%	13x
Research	Earlier than Legal/Finance	Not separately stated	56x
Customer Support	Not separately stated	Not separately stated	32x

How Agents Actually Work Differently to Chatbots

The distinction matters and is worth stating precisely. A chatbot interaction is short and self-contained: you ask, it answers, you move on. An agent operates independently for minutes or hours, orchestrating tool calls, interacting with environments, and iterating towards a solution without needing you to hold its hand.

OpenAI's paper estimates task horizons using an LLM-as-judge methodology. By May 2026, 70.2% of sampled individual Codex users had made at least one request estimated to exceed one hour of human work. A quarter had made a request estimated to exceed eight hours. At the 99th percentile of daily active users inside OpenAI, individuals were generating more than 60 hours of Codex agent turns in a single day, running multiple parallel agents simultaneously.

That last figure deserves a moment. One person, coordinating the equivalent of more than 60 person-hours of work, in a day. The unit economics of knowledge work have changed.

The Non-Developer Shift Is the Part Worth Acting On

Codex started as a coding tool. That framing is now obsolete. Non-developer individual users grew 137x between August 2025 and June 2026. Non-developer organisational users grew 189x. OpenAI's own non-developer employees, who started from a higher base, still grew 12x.

More telling is what non-developers are actually doing with it. More than a quarter of Codex output from business function employees, Finance, Marketing, Operations, was classified as engineering or coding work. Workers are crossing task boundaries that previously required a specialist or a ticket to the engineering queue. That is not efficiency. That is a change in what a single headcount can produce.

We have been advising clients on how agentic AI shifts content and search workflows for some time, and this data confirms what we have seen directionally: the non-technical adoption curve is steeper than most leadership teams expect, and it arrives faster than anyone plans for.

What This Means for Your Business Right Now

The practical question for a founder, CMO, or VP of Growth is not "should we adopt AI agents" but "which workflows do we redesign first, and who owns that decision." OpenAI's paper is useful precisely because it shows the sequence: engineering moves first, then functions with high task complexity and measurable output, then everyone else.

If your competitors' teams start using agents to complete in hours what previously took days, and your team is still treating AI as a chatbot assistant, the gap compounds quickly. It does not show up in a single quarter. It shows up in velocity, in hiring leverage, and eventually in the quality of decisions made with more or less context.

The organisations in the research that adopted agents earliest were not those with the most technical staff. They were those with low friction access to capable tools and a culture that permitted experimentation across role boundaries. That is a leadership and process question, not a technology one.

If you want to understand how agentic AI intersects with how buyers discover and evaluate B2B SaaS products today, our GEO advisory work is directly relevant: agents are increasingly the interface through which your prospects find and assess vendors, and that changes what good discovery content looks like.

The concrete action: map your three highest-volume, cross-functional workflows and ask honestly whether an agent could handle the execution, not the thinking, this quarter. Start there. The data from OpenAI suggests that once one function crosses over, the rest follow quickly.