OpenAI Signals: ChatGPT Work Use Across 165 Activities and ~100 Countries

OpenAI released a new slice of its Signals data on October 26, 2026, and it is built for researchers rather than headlines. According to OpenAI's announcement, the dataset links how consumers use ChatGPT for work to 165 work-activity categories across nearly 100 countries, reported quarterly from the first quarter of 2025 through the third quarter of 2026.

The one headline figure: across the 74 countries observed in every quarter, the ten most common work activities represented an average of 51.1% of each country's classified work messages in Q3 2026, down from 56.0% in Q1 2025. OpenAI's reading is that work use is spreading out, and that different countries are specializing in different kinds of work.

TL;DR

What is actually in the dataset?

OpenAI Signals is the company's ongoing effort to publish evidence about how people use AI. The Signals hub lists an individual-usage dataset, an enterprise-usage dataset and a run of reports, including pieces on how ChatGPT adoption broadened in early 2026, how workers use AI for tasks outside their usual occupation, and the shift to agentic AI seen in Codex usage.

This release adds a work-activity lens. Instead of asking only "how many people use ChatGPT," it classifies work-related messages into IWAs, which come from O*NET, the standardized framework that groups occupation-specific tasks into broader activities shared across jobs. Examples OpenAI names: preparing instructional materials, developing marketing content, programming and analyzing business data.

Because IWAs are a common language in labor economics, researchers can line the ChatGPT numbers up against employment and occupation datasets. The citation OpenAI gives for the release is "OpenAI Signals v3.0," authored by its Economic Research team.

How this release builds on earlier Signals reports

The Signals hub shows a steady cadence through 2026, which helps place this release. In May, OpenAI published pieces on how frontier firms are pulling ahead and how ChatGPT adoption broadened in early 2026, citing growth among users over 35. June brought a map of near-term AI job impacts in the EU labor market, a report on global adoption, and a Codex analysis of agentic work. July looked at how occupation-specific tasks spread across job boundaries. August covered how ChatGPT Enterprise spreads inside organizations and a randomized trial on students. September examined workers using AI for tasks outside their typical occupation.

Each earlier report answered one question with one slice of data. The October 26 release is different in kind: it publishes the underlying activity-level table so that outsiders can ask their own questions, using the O*NET occupational framework from the O*NET resource center as the shared vocabulary. That is a transparency step worth noting, even with the exclusions listed below.

What does "less concentrated" mean?

The top-ten share dropping from 56.0% to 51.1% means a smaller fraction of work messages falls into the most popular activities. Two readings are possible, and OpenAI mentions both. Countries may be specializing in the work they actually do, so a country strong in one industry shows up with a different mix. Or access to AI tools varies. OpenAI says these differences "might reflect variation in the kind of work done across countries or varying access to AI tools."

What the data cannot say on its own is causation. A broader spread could also reflect new users with different habits arriving, changes in the product, or changes in how messages are classified. OpenAI frames the release as a starting point for those questions, not as a finding about them.

ChatGPT usage vs the US economy

The most interesting chart in the release compares the distribution of ChatGPT work messages with the distribution of activities performed across US jobs. OpenAI says the two distributions differ substantially, largely because about 43% of the employment-weighted task mix for mapped US occupations falls into an "Other IWAs" bucket outside the individually reported categories. Those include moving materials, operating vehicles, and maintaining health or medical records. They did not meet the threshold for consistent representation in the original US ChatGPT sample.

Read that carefully. It does not say AI is useless for those jobs. It says people do not bring that kind of work to a chat window in large enough numbers to measure, which fits the idea of a digital-first, desk-work skew. It is a useful guard against the claim that chat usage maps neatly onto the labor market.

What the data leaves out

• Enterprise and Codex are excluded. OpenAI says so directly. Much of the coding and agent-driven work that has moved to Codex and enterprise plans is therefore outside this slice. For that side of the picture, see our coverage of how OpenAI's agent products are priced and limited.
• Work-related messages only, on individual plans. Consumer use for work is a mix of personal and professional behavior that varies by country and employer policy.
• Classification is automated. Messages are labeled by automated systems after identifying information is removed. Errors in classification propagate into the shares.
• Small groups are withheld and noise is added, which protects privacy but means very small country-activity cells are not reliable.
• It is OpenAI's own data and OpenAI's own methodology. Independent replication is hard because the underlying messages are not public. The downloadable aggregates are what researchers get.

How this fits the wider "AI and work" picture

Two threads from this month are useful context. A new benchmark effort argued that about 90% of US working hours are not covered by current AI benchmarks, which is the benchmark-side version of the same gap OpenAI shows on the usage side. And Anthropic's S-1 analysis touches the other major lab's claims about workplace adoption, with a different data source and a different incentive.

Comparing lab-published usage data is worth doing with discipline. Each lab picks its own categories, populations and time windows. OpenAI's choice to use O*NET categories is the part that makes this one easier to cross-check against government statistics. Anthropic and others use different taxonomies, so shares are not directly comparable.

How to use the data

1. Start with the interactive chart. Pick a country, drill from broad activity groups into specific IWAs, and watch how the mix changes over the seven quarters.
2. Download the files from the Signals data page and join them to O*NET and national employment tables on the IWA identifier.
3. Look at concentration, not just rank. The top-ten share is a single scalar you can compute per country and quarter to see who is specializing.
4. Check the cell sizes. Treat small countries and rare activities with caution given the suppression and noise.
5. State the exclusions in anything you publish. Enterprise and Codex are out; this is consumer-plan work messaging.

What this means for people building with AI

For builders and learners, the practical takeaway is modest but real. The activities where people bring the most work to ChatGPT tell you where demand for AI-assisted tools already exists, and the flattening of the distribution suggests long-tail work categories are growing. If you build tools for a niche activity, this data is a free way to size how much people already do it through a general chatbot, and what a specialized product has to beat.

For anyone writing about AI and jobs, the release is a reminder to separate three things: what people ask AI to do, what jobs consist of, and what AI can do reliably. This dataset speaks to the first, compares it to the second, and says nothing about the third.

Frequently asked follow-ups

Is ChatGPT use for work growing? The release is about composition, not volume. The headline number is a concentration measure, so it does not by itself say usage rose or fell. OpenAI has published separate adoption reports on the Signals hub.

Can I see my own country? The interactive chart lets you select a country and view its mix over time, subject to OpenAI's small-group suppression.

Does this show which jobs AI will replace? No. It shows which work activities people bring to ChatGPT. It does not measure task completion quality or substitution.

Related reading on explainx.ai:

• Mercor benchmark map: 90% of US working hours not covered by AI benchmarks
• Anthropic S-1 explained, including workplace adoption claims
• OpenAI Dots Pro $200 usage halved
• AI readiness checklist for business leaders
• Why AI companies want you using agents: token economics
• OpenAI Dots always-on agents at DevDay

Figures are as published by OpenAI on October 26, 2026 and may be revised in later data versions.