OpenAI Says Agents Have Replaced Chatbots as Its Default Work Tool

Codex now accounts for 99.8% of weekly output tokens generated inside OpenAI, and the average OpenAI employee spends more than 85% of their output tokens on the agent — a reversal from August 2025, when the average OpenAI worker spent less than 10% of their tokens on the tool.

Those figures come from OpenAI's own account of a year of internal and external Codex usage, published in a report titled "How agents are transforming work." The company frames the shift as evidence of a structural change in how knowledge workers use AI: agentic tools, which can operate independently for minutes or hours while orchestrating tool calls and iterating toward solutions, are replacing the short, self-contained interactions that define chatbots.

The stakes extend beyond OpenAI's walls. The company argues its data shows "what unfolds when people have broad, low-friction access to capable agentic tools: as the tools improve, people use them for longer, more complex, and more cross-functional work." Businesses redesigning workflows, employees deciding which skills to invest in, and policymakers studying AI's effect on the labor market all have a direct interest in whether that pattern holds beyond one company's workforce.

Longer tasks, not more chats

The clearest signal in OpenAI's data is the lengthening horizon of delegated work. By May 2026, 80.6% of sampled individual users had made at least one Codex request estimated to exceed 30 minutes of human work. 70.2% had made a request exceeding one hour. 25.6% had submitted at least one request estimated to exceed eight hours of human work.

Growth was fastest in the longest-duration bucket, though from a low base. From December 2025 to May 2026, the share of users requesting work equivalent to more than one hour of human effort climbed to 70.2%, and the share requesting more than 30 minutes reached 80.6%. In May 2026, more than 70% of users asked Codex to complete a task that would take a person more than one hour.

Heavy users push that horizon even further. Among daily active users at OpenAI, users at the 99th percentile regularly generated more than 60 hours of Codex agent turns per day by June 2026, distributed across multiple parallel agents. OpenAI describes this as a behavioral shift from asking Codex for one answer at a time to "increasingly orchestrating multiple agent tasks over the course of a day."

That orchestration pattern matters for capacity planning and economics. Sixty hours of agent runtime per person per day is only possible with parallel execution, and it changes the unit of consumption from a conversation to something closer to delegated labor — with corresponding implications for compute demand and token costs.

From engineering to legal and recruiting

Adoption inside OpenAI followed a predictable path and then accelerated in unexpected places. Engineers began using Codex first, gradually. The average engineer at the company shifted the majority of their OpenAI product usage to Codex by December 2025. Today, the average engineer generates 99% of their output tokens with Codex rather than ChatGPT.

Non-technical departments moved later but faster. Legal, Finance, and Recruiting crossed into majority Codex use around April 2026. The average lawyer or recruiter at OpenAI now generates more than 85% of their output tokens on Codex, according to the report. Every department, including non-technical ones, now uses Codex as its primary AI tool for work.

Usage has also intensified sharply over the past six months. Among active internal users, median combined output tokens rose 56 times in Research between November 2025 and June 2026 — the biggest jump of any department. Customer Support rose 32 times and Engineering rose 27 times. Legal grew more gradually but still reached 13 times its November 2025 level.

Non-developers are the fastest-growing group

Across individual users, organizational users, and OpenAI workers, Codex began as a developer tool — the natural audience for a coding agent. As the product expanded toward general knowledge work, non-developer adoption grew even faster than developer adoption.

The numbers are striking. Since August 2025, non-developer individual users multiplied 137-fold by early June 2026. Non-developer organizational users increased 189-fold. Within OpenAI, non-developer users rose 12-fold, which OpenAI attributes to a starting point that was already "well above average."

The company adds a caveat: the shift does not mean every non-developer uses Codex the way an engineer does. It means more non-developers are using Codex for some kind of agentic work.

That distinction shapes how to read the occupational data. OpenAI compared inferred occupations within the company against the type of work appearing in Codex outputs. Engineering and coding remain the largest output category for data science and research staff, while knowledge work dominates for finance and business operations, marketing, operations, and other departments.

But the boundaries are blurring in one direction in particular: more than one-fourth of the work done with Codex by workers in business functions was engineering or coding. OpenAI says agents "can lower the cost of moving across task boundaries and help workers do adjacent work that used to require more specialized technical support." Non-technical users regularly deploy Codex for automation, data transformation, tooling, debugging, and structured analysis.

Why the internal numbers matter externally

OpenAI is reporting on its own product, using its own employees as the most visible cohort — a limitation worth keeping in mind. OpenAI workers sit at the frontier of AI familiarity and have free, low-friction access to the company's most capable tools. The report itself acknowledges this framing: it presents frontier users adopting capable agentic tools at the frontier.

Even with that caveat, the trajectory is concrete and quantified. Codex went from under 10% of the average OpenAI worker's tokens in August 2025 to a dominant share within roughly a year. Task horizons stretched from minutes to multi-hour delegations. Non-developer populations grew by two orders of magnitude. And the heaviest internal users now run the equivalent of multiple full-time agents in parallel every day.

The report closes with a forward-looking claim grounded in that data: "As time goes on, this is likely to be what the future of work looks like." For organizations tracking whether agentic AI shifts from a developer convenience to a company-wide replacement for chatbot interaction, OpenAI's internal crossover — non-technical departments crossing over in weeks, not years — is the data point to watch.