OpenAI paper says Codex non-developer use grew 137-fold; lawyers and recruiters now route 85% of output through agents

A working paper from OpenAI's Economic Research team, published June 25, documents a rapid shift from chat to delegated agentic work — though every figure is drawn from OpenAI's own platform telemetry.

OpenAI’s Economic Research team published a working paper on June 25 claiming that non-developer use of its Codex agent has grown 137-fold among individuals and 189-fold inside organizations since August 2025, with lawyers, finance staff, and recruiters at OpenAI itself now routing more than 85% of their AI output through Codex rather than ChatGPT. The paper, titled “The Shift to Agentic AI: Evidence from Codex,” is the most detailed picture yet of how delegated agent work is replacing turn-by-turn chat, and it’s also a marketing artifact for the product whose adoption it measures.

The internal numbers are the headline. As of June 11, Codex generated 99.8% of weekly output tokens produced by OpenAI employees. Inside the company, legal, finance, and recruiting crossed majority Codex usage around April 2026, earlier than engineering did. For paying organizational customers the figure was 63.3%; for individual users, 16.5%.

What the paper calls “an unusually favorable agentic environment” is OpenAI describing itself.

The more analytically interesting metric is task duration. By May, 80.6% of sampled individual users had submitted at least one request the model estimated would take a human over 30 minutes; 25.6% had submitted a request estimated at over eight hours. Delegation of 8-plus-hour tasks grew nearly tenfold in the first half of 2026. Codex turns invoke external tools 60.3% of the time versus 21.9% for ChatGPT, and over 10% of users run three or more concurrent agents weekly. Inside OpenAI, 99th-percentile daily users generate more than 60 hours of agent turns per day across parallel sessions.

External signal exists. Samsung Electronics’ South Korean Codex base grew nearly 800% since February 1, per TechTimes, spanning marketing, manufacturing, and product-design staff with no programming background. The Next Web puts total Codex weekly users above 5 million, roughly 20% of them non-developers.

The caveats sit in the footnotes. Axios notes that the individual-user sample is drawn from the 0.1% of users who opted in to allow queries for training, and that human-equivalent durations are model-estimated. The Next Web flags that every figure is OpenAI’s own telemetry, self-reported and unaudited. The 2008 framing of bank stress tests was also produced by interested parties; what mattered then, and matters now, is whether anyone outside the firm can replicate the numbers.

OpenAI paper says Codex non-developer use grew 137-fold; lawyers and recruiters now route 85% of output through agents

Sources