Codex becomes an observatory of agentic work.

OpenAI uses Codex to measure agentic work, revealing that 80.6% of users delegate tasks over 30 minutes and internal employee adoption exceeds 85% of tokens.

In a new Economic Research paper, OpenAI uses Codex to measure the shift from short chat to long-form task delegation. The focus is no longer just code: the company describes a use case where the agent can work for several minutes or several hours, call tools, interact with an environment, and progress through iterations. According to OpenAI, 80.6% of a sample of individual users entrusted Codex with at least one task estimated at over 30 minutes of human work, 70.2% a task estimated at over an hour, and 25.6% a task estimated at over eight hours. These thresholds should still be read as orders of magnitude, as they are estimated by a model based on conversations.

The other interesting part comes from internal use: Codex has reportedly become the primary AI tool across all OpenAI departments, including Legal, Finance, and Recruiting. The company indicates that Codex's share exceeds 85% of output tokens for the average employee, and reaches 99.8% of weekly output tokens generated internally. OpenAI particularly emphasizes the progress of non-developer profiles, with adoption reportedly increasing faster than among developers. Implicitly, the paper tells less about the rise of a coding tool than the advent of a new unit of work: no longer an answer, but an entire task entrusted to an agent, sometimes in parallel, sometimes outside the usual professional scope.