ChatGPT’s top 10 work tasks slipped from 56.0% to 51.1%

OpenAI tracked work use of ChatGPT in 74 countries in every quarter. In Q3 2026 the ten most common work activities made up an average of 51.1% of each country’s classified work messages, down from 56.0% in Q1 2025, so consumer work use is now less concentrated in the usual top tasks. The data covers individual ChatGPT plans only, with no enterprise messages and no Codex use.

OpenAI published the figures on October 26 as a new Signals data release. It maps consumer work use of ChatGPT to 165 activity categories across nearly 100 countries, quarter by quarter from Q1 2025 through Q3 2026. The categories are Intermediate Work Activities (IWAs) from O*NET, a standard framework that groups the tasks specific to each occupation into broader activities shared across jobs. OpenAI’s examples run from preparing instructional materials and developing marketing content to programming and analyzing business data. Because the framework is common, researchers can set ChatGPT use against other occupational and employment data, follow it over time and match it to other economic indicators.

The 51.1% is an average of country-level shares, taken over the 74 countries present in every quarter. It is not a global share of all messages. OpenAI reads the decline as countries specializing over time, with different countries moving into different types of work. It says the differences might reflect the kind of work done in each country or varying access to AI tools, and offers that as one question researchers can pursue with the data.

The release also compares the ChatGPT mix with the activities performed across U.S. jobs, and the two distributions differ substantially. Most of that gap comes from one bucket. About 43% of the employment-weighted task mix for mapped U.S. occupations falls into “Other IWAs”, which covers activities such as moving materials, operating vehicles and maintaining health or medical records. Those categories did not meet OpenAI’s threshold for consistent representation in the original U.S. ChatGPT sample, so they have no individual ChatGPT figure to set against the U.S. economy.

Researchers do not read individual conversations. Automated systems classify messages after identifying information is removed. OpenAI publishes only aggregated statistics, withholds results for small groups and adds statistical noise to protect individual messages.

The earlier baseline is “How People Use ChatGPT”, an NBER working paper by OpenAI’s Economic Research team and Harvard economist David Deming that came out in September 2025. It found steady growth in work-related messages but faster growth in non-work messages, which went from 53% to more than 70% of all usage. On average, writing made up 40% of work-related messages in June 2025. That study sorted conversations into broad topics such as “Writing” and “Practical Guidance” rather than IWAs, so its shares do not map one-to-one onto the new categories. CNBC reported at the time that the paper had not been peer reviewed.

Programming is one of the activities in the new data. The Harvard Business School AI Institute put computer programming at 4.2% of messages across all ChatGPT users as of June 2025. Because Codex is excluded, whatever programming share the new release shows covers chat on individual plans only.