Epoch AI, a research organization, launched a data browser called the ChatGPT usage explorer on October 1. The data comes from pollster YouGov, which randomly selected 5,000 ChatGPT users from a US panel and asked them to export and share their own chat histories, spanning ChatGPT's launch in November 2022 through December 2025 — about 660,000 conversations and 8.3 million messages in total. The exports took place between January and August 2026.
What sets this data apart from earlier surveys is that it comes from actual usage records rather than respondents' recollections or statistics released by OpenAI itself. The trade-off is that no content is visible: YouGov stripped out all message text before handing the data to Epoch, so Epoch's researchers never read a single conversation. That means the explorer can't sort usage by topic or purpose — only by frequency, length and timing metadata.
Usage keeps climbing
Several numbers point the same way. The median number of messages an active user sends per month rose from 14 in January 2023 to 36 in December 2025; the mean rose from 31 to 158.
The shift in how many days people use it per month is even more pronounced. Between December 2023 and December 2025:
- The share of users active just 1 day a month fell from 32.6% to 22.1%;
- The share active 11 to 20 days a month rose from 8.1% to 17.0%;
- The share active 21 or more days a month rose from 2.6% to 10.5% — roughly quadrupling.
The median length of a single conversation has stayed flat at 2 prompts, while the mean length rose from 3.7 to 5.8. The median number of conversations an active user starts per month rose from 3 to 5.
A small group accounts for most of the activity
Epoch's other headline figure: the most active 10% of users sent 63% of all prompts.
Running a rough estimate on the totals: the 8.3 million messages include model replies, and assuming user prompts and model replies split roughly 50-50, users sent around 4.15 million prompts. The top 500 users account for 63% of that — about 2.6 million prompts, or roughly 5,200 per person, close to five a day on average over three years. The remaining 4,500 users sent about 1.55 million prompts combined, roughly 340 per person, or about one every three days. This is only an estimate — Epoch did not publish an exact total prompt count.
The gap between the mean and the median tells the same story: that monthly average of 158 messages is driven mainly by a small group of heavy users.
Where the sample falls short
Epoch spelled out the limitations clearly. The biggest is survivorship bias: the sample only includes people still using ChatGPT in 2026, leaving out anyone who tried it between 2022 and 2025 and then stopped. At least part of the "climbing usage" trend above is an artifact of that selection.
The demographics skew from the broader population too. College graduates make up 55% of the sample versus 40% of US adults overall; Hispanic respondents make up 8% versus 19% overall. The sample also skews older than ChatGPT's overall user base: those under 30 make up 14% of the sample versus 29% overall, while those 65 and older make up 18% versus 10% overall.
Several questions remain open. The metadata records which model generated each reply and which tools it called, but the explorer does not publish a breakdown of model usage; whether participants were paid for taking part isn't stated and can't currently be verified. OpenAI has not publicly responded to the dataset so far. Epoch has made breakdowns by gender, age, income, education, employment status and subscription type available in the downloadable dataset.
Sources: Epoch AI's official release, CocoLoop, YouGov sample notes; figures cross-checked against Epoch's methodology notes for sample size, total messages and demographic ratios.