OpenAI says 251 scientists in its ChatGPT for Academic Researchers pilot estimated that their work would have taken 1.9 times as long without AI. The figure comes from a survey, and OpenAI’s own study, published on 16 November, says “Productivity is self-reported” and that its findings “do not establish causal effects”.
The usage side of the study is larger: OpenAI’s Economic Research team classified 580,186 scientific messages from 1,819 scientists sent between 20 August and 22 September. Our number: that is about 319 scientific messages per scientist over the 34-day window, or about 9.4 a day on average, by our arithmetic.
Update, 16 November 2026: OpenAI said in an X post at 17:13 UTC that it is “expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support”. Applications open on 23 November, and OpenAI says: “By default, we don’t use data shared through the program to train our models”.
The post quotes OpenAI’s 29 July launch post, which set access “starting with 10,000 researchers and expanding to 100,000 through 2027”. The 200,000 goal is double that 100,000, as OpenAI’s program post also says, and 20x the 10,000 starting point, by our arithmetic.
Business Pill · Concentration explained
An evergreen Business Engineer explainer on concentration, when a few big pieces make the total. It teaches the general idea only and does not discuss OpenAI’s study.
The key insight: As we read OpenAI’s study, its two halves measure different things: classifiers counted 580,186 scientific messages from 1,819 scientists, about 319 each by our arithmetic, while the 1.9x productivity figure is the estimate of 251 survey respondents, which OpenAI itself labels self-reported.
What OpenAI Measured
OpenAI’s blog post and the 76-page paper by Chris Liao, Carl Shan and Aaron Chatterji rest on two datasets. The first is usage: the 1,819 scientists sent 1,262,893 messages in all, of which 580,186 were classified as scientific or scholarly, the paper says.
OpenAI says the research team “did not read the underlying messages or ChatGPT’s responses”. Automated classifiers sorted each message by activity, and the team saw only those categories. Reading and learning took 28% of the average scientist’s scientific messages and writing and communicating 27%, with each scientist weighted equally.
Use was uneven. The paper says “the most active 10% of scientists account for 42% of scientific-message activity”, and the median number of scientific messages on a scientist-day was 2. Our mean of about 9.4 a day sits well above that median; the paper calls the distribution right-skewed.


What Was Posted on X
OpenAI posted at 17:13 UTC on 16 November that it is expanding free access to its frontier models for up to 200,000 scientists, with applications opening on 23 November. The post quotes OpenAI’s 29 July launch post, which set a path from 10,000 to 100,000 researchers through 2027; this piece uses it for the program figures in the update box.
Greg Brockman, OpenAI’s president, posted at 17:29 UTC on 16 November, quoting OpenAI’s own post on the study with the words ‘ChatGPT for doubling the productivity of scientists’. This piece quotes his line and OpenAI’s ‘nearly doubling’ summary as their wording; the study’s 1.9x figure is the self-reported estimate of 251 surveyed scientists, as OpenAI’s blog post describes it.
Where the 1.9x Comes From
The second dataset is a survey run from 30 September to 7 October. Out of 270 valid completed responses, 251 scientists consented to inclusion, the paper says. They were asked how much more time they would have needed without AI for the same work at the same quality.
OpenAI says that across eight research activities, “one hour of work with AI available would have taken about 1.9 hours without it”. The paper reports mean ratios from 1.35x in administrative work to 2.83x in research software and code development. Answers were picked from set bands, with “400% or more” additional time coded as 5.
The paper flags the sample itself: “respondents are a selected group of scientists who applied for expanded AI access and chose to answer a survey about AI use”. OpenAI’s X account summarised the result as scientists who “reported nearly doubling their productivity”; Greg Brockman quoted it as “ChatGPT for doubling the productivity of scientists”.
Who Asks the Research Question
OpenAI says 68% of scientists used ChatGPT to develop research ideas, defined as sending at least three ideation messages. In those messages, scientists brought a specific question 35% of the time and asked ChatGPT to refine their idea 36% of the time; 23% of the time they “asked ChatGPT to suggest a research question”.
Scientists followed up to clarify, check or revise an answer after 26% to 38% of requests, depending on the activity, the blog post’s chart shows. Asked about limits, 51.8% named time or staffing and 43.8% funding, and 6.4% reported no major bottlenecks, the paper says.
From Pro to Plus
The study came with an expansion of the program. OpenAI says 20,000 researchers in 33 countries have joined since July, and it will award 50,000 new grants, each covering the recipient and up to two collaborators, with applications opening on Monday, 23 November.
The pilot gave researchers ChatGPT Pro. OpenAI says that when it reviewed pilot usage, “most participants would have stayed within ChatGPT Plus usage limits”, so new grants get Plus. Its goal is 200,000 researchers, 10x the 20,000 who have joined so far, by our arithmetic.
Our earlier pieces covered the first post on OpenAI’s science blog, an OpenAI physics test and the Claude Gödel-theorem results, which were about what models solved. This study is about how scientists use the product day to day.
The Structural Read
As we read the usage data, ChatGPT is mostly a reading and writing tool for these scientists: those two activities take 55% of the average scientist’s scientific messages by our arithmetic, against 5% for developing ideas.
As we read the program post, the move from Pro to Plus is a cost decision OpenAI drew from the same usage data: it says most pilot participants would have stayed within Plus limits, and it is widening access to 50,000 new grants.
OpenAI, Science at the Frontier
“Understanding whether those expectations translate into faster discovery will take more research.”
Three Implications
FOR RESEARCH LEADERS The paper reports the largest self-reported gain, 2.83x, in research software and code development, and the smallest, 1.35x, in administrative work.
FOR READING THE 1.9X The paper says respondents applied for expanded AI access and chose to answer the survey, so the figure may not carry over to scientists in general.
RELATED Our read of OpenAI’s science blog covers its first post.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Subsidized AGI Economy.
The framework argues: “The negative margin is not a leak. It is a procurement budget.”
As we read it, OpenAI’s free academic program fits that frame: OpenAI gave the pilot free Pro access, then used what it learned from pilot usage to move new grants to Plus and reach more researchers. The framework predates the study and does not discuss it.
What Is Not Established
The 1.9x is the survey respondents’ own estimate, not a measured output, and OpenAI says the findings “do not establish causal effects”. The discovery figures are expectations: 98% of the 250 who answered expected AI to speed discovery in their field.
We read the blog posts in full and the paper’s abstract, introduction, data, survey and results sections and its program appendix. We did not contact OpenAI or the authors for this piece.
The Bottom Line
OpenAI classified 580,186 scientific messages from 1,819 scientists, about 319 each by our arithmetic, and 251 surveyed scientists estimated their work would take 1.9 times as long without AI. OpenAI labels that figure self-reported and its findings non-causal.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The study figures come from OpenAI’s blog post “Science at the Frontier: How scientists use ChatGPT” of 16 November 2026, including the labels on its charts, and from the accompanying paper by Chris Liao, Carl Shan and Aaron Chatterji, of which we read the abstract, introduction, data and survey sections, results sections 4.2 to 5 and the program appendix. Program figures are from OpenAI’s post on broadening access, also of 16 November.
The usage figures come from automated classifiers, as OpenAI describes; the productivity figure is self-reported by survey respondents. Messages per scientist, the daily mean and the 10x goal ratio are our arithmetic. OpenAI and the authors were not contacted for this piece.
Sources: OpenAI, Science at the Frontier: How scientists use ChatGPT, 16 Nov 2026 · Liao, Shan and Chatterji, Science at the Frontier: How Scientists Use ChatGPT (paper) · OpenAI, Broadening access to AI for scientists, 16 Nov 2026 · OpenAI on X, 16 Nov 2026 · Greg Brockman on X, 16 Nov 2026









