OpenAI has released a paper claiming that its staff have switched almost entirely from its ChatGPT chatbot to its agentic Codex platform in what it said “reflects what we believe will be the future of work”.
97.9 per cent of OpenAI employees now use Codex, and it makes up 99.8 per cent of the output tokens for the organisation. The discrepancy between the numbers reflects the higher token usage that comes with agentic AI, the company said.
The usage rates for active OpenAI product users have also risen, with 17.3 per cent of them
now using Codex, up from around 1 per cent in August 20025. For individuals, the numbers are far lower, with only 0.7 per cent of active users opting for Codex instead of or as well as ChatGPT.
It is here that the higher token usage becomes more apparent, with OpenAI reporting that Codex makes up 63.3 per cent of output tokens for organisations, and 16.5 per cent for individuals.
This means that for organisations, Codex produces around 3.7 times as many tokens per user as ChatGPT, while for individuals this number reaches a multiple of 23.6.
This is in part because Codex users ask the agents to perform longer tasks. OpenAI said that nearly a quarter of all Codex requests are for tasks that would take a person more than an hour to complete based on a large language model’s judgement.
Internally, token use has skyrocketed, which OpenAI attributes to Codex usage having “deepened and intensified”. Between November 2025 and June 2026, token use across departments rose sharply. Those in its research division increased token output by 53 times, customer support 32 times, and engineering 26 times.
In the first quarter of 2026, over 25 per cent of users made at least one Codex request that was estimated to take a human over eight hours.
By June, the 99th percentile of Codex users were asking Codex to run more than 60 hours of agent turns a day over multiple agents, with the average user generating fewer than 10.
The company describes this as demonstrating “how Codex has transformed how OpenAI uses AI to do productive work,” later adding that “as the tools improve, people use them for longer, more complex, and more cross-functional work”.
The company did not provide statistics on productivity or return-on-investment changes based on this token use.
OpenAI is currently on track to lose $14 billion this year based on its own projections and is considering significant price cuts in anticipation of a fight for users with rival Anthropic, according to the Wall Street Journal.
It is widely reported that the company’s current subscription models allow users to spend tokens that would cost orders of magnitude more on a per-token basis. Critics including author Ed Zitron have noted that OpenAI’s costs appear to scale linearly with its revenues, meaning that the cost equation does not become more favourable to it as token usage increases.
