OpenAI published quarterly ChatGPT work-use data spanning 165 activities and nearly 100 countries, covering the first quarter of 2025 through Q3 2026.
OpenAI Economic Research released a dataset on October 26, 2026, intended to help researchers study how people use ChatGPT for work. It classifies usage across 165 work activities in nearly 100 countries, with quarterly data from the first quarter of 2025 through the third quarter of 2026.
The activity categories are based on O*NET, a framework that groups occupation-specific tasks into broader types of work. OpenAI says the common classification makes it possible to compare ChatGPT usage across countries and over time, and to connect the patterns with other economic indicators. Examples of the activities represented include preparing instructional materials, creating marketing content, programming and analyzing business data.
One reported pattern is a decline in the share of messages concentrated in the most common activities. Across the 74 countries represented in every quarter, the ten most frequent activities accounted for an average of 51.1% of each country’s classified work messages in the third quarter of 2026, compared with 56.0% in the first quarter of 2025. OpenAI says this may reflect different mixes of work or differences in access to AI tools; the dataset is intended to let researchers investigate such explanations, not settle them.
The release also allows comparisons between ChatGPT use and the employment-weighted mix of tasks across U.S. jobs. OpenAI notes that about 43% of that mapped U.S. task mix falls into “Other” activity categories, including moving materials, operating vehicles and maintaining health or medical records. Those activities were not individually reported because they did not appear consistently enough in the underlying U.S. ChatGPT sample.
The data covers work-related messages from individual ChatGPT plans; it excludes enterprise messages and Codex use. OpenAI says researchers do not read individual conversations: automated systems classify messages after identifying information is removed, and published results are aggregated, with small groups withheld and statistical noise added for privacy.