Codex is quietly eating ChatGPT’s lunch, and even non-developers are adopting agents quicker than engineers did, with skills emerging as a way to standardise how teams actually use AI.

This is according to OpenAI’s new research, the most useful internal dataset on agentic adoption anyone’s likely to get this year. Through August 2025, the average OpenAI worker spent less than 10% of their tokens on Codex. By June 2026, Codex accounts for 99.8% of weekly output tokens generated inside the company. That’s not just engineers switching tools. That’s the entire organisation, including Legal, Finance, and Recruiting, moving from chat to delegation as the default way of working.

Generative AI is maturing and changing work patterns fast. The chatbots we know and love answer questions. Agentic tools do the work. OpenAI’s paper draws a clean line between the two. Once a tool starts producing rather than just advising, the unit of work stops being a conversation and starts being a delegated task that runs for minutes, hours, sometimes most of a day.

The non-developer numbers back this up hard. Since August 2025, non-developer users of Codex grew 137x among individual accounts, 189x among organisational accounts, and 12x inside OpenAI itself. Lawyers and recruiters at the company now generate more than 85% of their output tokens through Codex, a coding tool that’s clearly stopped being just a coding tool.

So what are they all doing? The report doesn’t explicitly say it but the ability to create whole working dashboards, document designs, wireframes etc, has become significantly demoncratised by agentic coding and MCPs accessing other tools.

Over a quarter of the work business-function staff did with Codex was engineering or coding, tasks they had never have touched without an agent doing the heavy lifting. Agents are lowering the cost of crossing into adjacent skills, and the job description on file is increasingly disconnected from the work actually getting done.

Skills get one quiet mention in the research too. It notes that intensive users are more likely to use skills, run longer tasks, and operate several agents in parallel, rather than firing off one-off requests. It’s an early signal that skills might become the standard unit firms use to package and scale how employees work with AI, rather than every person improvising their own prompts.

One thing the paper’s own framing slightly undersells is that output tokens are a clean metric to report, but they are actually a shaky proxy for what’s really happening under the hood. A single Codex session dragging in long file context, running multiple iterations, and chaining tool calls can burn enormous compute while producing a modest output token count. User numbers and token shares make for a tidy chart. They don’t tell you how much work, or how much GPU time, is actually going into getting there.

OpenAI is upfront that its own workforce is an unusually generous test bed, cheap usage, high buy-in, workflows close to the product itself. Most companies won’t see adoption this fast. But the direction is a strong signal of the changing face of generative AI. Chat is becoming the legacy interface, and handing off whole tasks to agentic coding is becoming the default. But with millions of generative AI users still using nothing more than the odd basic chat for information or low level content, the skills gap will widen rather than close up.