Independent reporting on artificial intelligence.


The Frontier Wire

Research

How scientists use ChatGPT: reading OpenAI's Science at the Frontier data

On 16 November 2026 OpenAI published a usage study of 580,186 scientific messages from its academic researcher programme, alongside a bigger free-access offer and, a day earlier, a new science blog. The paper is careful about its limits. The headline numbers travel further than those limits do.

OpenAI’s new study of how scientists use ChatGPT is a detailed picture of one group: 1,819 academic scientists who applied for free access to OpenAI’s models and were let in. Over 34 days in August and September 2026 they sent 580,186 messages that a classifier labelled as scientific work. Most of those messages were about reading and writing. About two-thirds of the scientists used ChatGPT at least three times to develop research ideas. About a third of their requests drew a follow-up asking ChatGPT to explain, check or fix what it had produced. Separately, 251 of them answered a survey, and their estimates are the source of the most quoted number, that the same work would have taken 1.9 times as long without AI. That number is a self-reported estimate from a self-selected group, and the paper says so.

The study, “Science at the Frontier: How Scientists Use ChatGPT,” was published on 16 November 2026 with a key-findings summary and a 76-page paper by Chris Liao, Carl Shan and Aaron Chatterji of OpenAI’s Economic Research team. The same day OpenAI expanded the programme that supplied the data, ChatGPT for Academic Researchers, toward a goal of 200,000 researchers. A day earlier it launched a blog for science. Read together, the three posts are a research finding, a product offer and a publishing channel, and the research comes from the company whose product it studies.

The Frontier Wire has not seen the underlying data, which is not public. This article reads the full paper, the key-findings PDF, the three OpenAI posts, the programme’s help-centre page and Anthropic’s comparable study of Claude, published on 5 November. Numbers are the paper’s unless the text says otherwise. Where a figure is our arithmetic, it is labelled.

TL;DR

  • The usage data covers 1,819 scientists in OpenAI’s academic programme over 34 days; the survey covers 251 of them. Neither sample is designed to represent scientists in general, and the paper says the survey findings “may therefore not generalize.”
  • Reading and writing account for over half of the average scientist’s scientific messages. Ideation is a small share of messages but used by many: 67.6% of scientists sent at least three ideation messages.
  • Scientists followed up to clarify, check or revise 31.8% of initial requests on average, more often for tasks needing advanced expertise.
  • The 1.9 times time saving and the 98% who expect faster discovery are survey answers, not measurements of output.
  • The programme expansion swaps the pilot’s ChatGPT Pro for Plus, drops the lottery for first-come access and opens on 23 November 2026.

Three posts in two days

The study did not arrive alone. The sequence matters because the programme that generated the data is also the product being offered.

DateEventSource
29 July 2026ChatGPT for Academic Researchers launchesPaper, Appendix A.1
10 August 2026Update announces a lottery for the initial 10,000 seats and a waitlistPaper, Appendix A.1
20 August to 22 September 2026Usage observation windowPaper, Section 2.1
30 September to 7 October 2026Survey fieldedPaper, Section 2.2
5 November 2026Anthropic publishes its study of how scientists use ClaudeAnthropic
6 November 2026Date on the Science at the Frontier paperPaper
15 November 2026OpenAI launches its science blogOpenAI
16 November 2026Study, key findings and programme expansion publishedOpenAI
23 November 2026Applications open for 50,000 new grantsOpenAI

The paper cites Anthropic’s report and compares its own findings with it, which makes the two studies the most useful check on each other. Both draw on the same task taxonomy, SciNet, and both classified usage with a language model rather than having people read conversations.

Who is in the data

The sample is defined by the programme. According to the paper’s appendix, ChatGPT for Academic Researchers targeted research faculty and postdoctoral researchers at eligible institutions in supported countries. Applicants had to verify their affiliation through SheerID with an institutional email address and provide a paper posted on arXiv, bioRxiv or ChemRxiv in the preceding three years, in one of seven fields: biological sciences, chemistry and materials, computer science, Earth and planetary sciences, engineering, mathematics, or physics. Seats were allocated by lottery among eligible applicants. Each approved applicant received a ChatGPT workspace and could invite up to four colleagues from the same institution. The paper notes that the published eligibility criteria “have not been verified for every recruitment period.” OpenAI’s programme post says the pilot gave participants ChatGPT Pro.

The cohort studied is the 1,819 scientists, counting workspace owners and collaborators, who sent at least one scientific message between 20 August and 2 September. That cut-off ensures each is observed for at least 21 calendar days. Together they sent 1,262,893 messages in the window, through ChatGPT’s Chat and Work modes and through Codex, including the Codex command line. The classifier labelled 580,186 of those messages as scientific or scholarly work. Another 237,034 could not be classified, which the paper reports in a footnote: scientific tasks are 45.9% of all recorded messages and 56.6% of classified ones.

Three features of this group shape every result. Its members chose to apply for expanded AI access. They were drawn from the seven fields, with no social sciences or humanities in the usage data. And they had free, high-limit access during the window, which could change how much and how they used it compared with a scientist paying for a plan.

Use within the group was uneven. The most active 10% of scientists accounted for 42% of scientific-message activity. On the median scientist-day the count was two scientific messages; the 75th percentile was 12 and the 25th percentile zero. To keep heavy users from dominating, the paper weights each scientist equally when it averages shares, a choice it applies throughout.

How the messages were classified

Nobody at OpenAI read the messages, the paper says. A model, GPT-5.6-Terra, labelled each one using the message and, when needed to resolve references such as “this” or “step four,” the immediately preceding exchange. The research team saw only the labels. The paper reproduces the classifier prompt, which instructs the model not to quote or summarise conversation text and not to infer the researcher’s identity.

Each scientific message received labels on two dimensions. The first is the kind of work: one of 10 workflows and 40 tasks, adapted from the SciNet taxonomy, with 20 of SciNet’s 27 task definitions kept verbatim, six reworded and 14 added. The second is four characteristics of how the scientist engaged with AI: whether the message is an initial request or a follow-up (and if so whether it seeks clarification, checking, revision, extension or a reaction); who supplied the research question in ideation messages; how long the task would take a competent person without AI; and the minimum expertise it requires.

Two choices in this design deserve attention. The duration and expertise labels are the classifier’s estimates, not measurements, following an earlier study of Codex use by Johnston and colleagues. And the paper, as The Frontier Wire read it, does not report a check of the classifier’s labels against human-coded examples, such as an agreement rate on a sample. Privacy constraints that keep researchers from reading messages make such a check harder, but they do not rule out validating the classifier on consented or synthetic data. Without it, readers cannot tell how often “Develop or test proofs” or “Formulate research problems” is assigned correctly, which matters most for the fine-grained ideation results.

The survey answers were also partly coded by a model. The open-ended question about how AI would speed discovery produced 224 written responses, which GPT-6 Astra split into 367 distinct claims and classified into mechanism categories.

What scientists used ChatGPT for

The headline usage result is concentration. Four workflows account for 85% of the average scientist’s scientific messages, and two of them, reading and writing, account for over half.

WorkflowAverage share of scientific messagesScientists with at least 3 messages in it
Reading and knowledge acquisition28%94.2%
Writing and communication27%88.5%
Scientific analysis (empirical and formal)17%79.0%
Research software and code14%61.7%
Ideation and hypothesis generation5%67.6%
Mentorship and teaching3%32.2%
Research administration3%42.2%
Peer review and service2%33.6%
Data gathering1%23.2%
End-to-end autonomous researchUnder 1%6.3%

Message shares are from OpenAI’s study page; uptake percentages from the paper’s Table A3.

At the task level, “Explain scientific concepts” is the single largest use, at 18.8% of the average scientist’s scientific messages, and 89.7% of scientists used it at least three times. “Draft manuscripts” is second at 13.9%. “Develop or test proofs” and “Create research visuals” each account for 4.2% of messages, but 26.2% of scientists used the first at least three times against 55.4% for the second. The paper’s reading is that AI “may be broadly useful for helping scientists learn and communicate,” while use of its technical capabilities “depend[s] more on scientists’ research methods.”

Field differences are real but mostly modest. Mathematicians have the highest share of scientific analysis (33.7%) and the lowest of research software (5.8%). Data gathering reaches 2.5% of messages in biology and in Earth and planetary sciences, against 0.2% to 1.0% elsewhere, which the authors link to large shared datasets such as sequencing databases and satellite archives. Younger scientists send more messages per day and spend a larger share on code: research software falls from 17.9% of messages among scientists aged 18 to 29 to 9.8% among those 40 and older. Ideation is the one workflow where the paper cannot reject equal average shares across all seven fields.

The comparison with Anthropic’s study is the most striking result in this section. OpenAI’s sample devotes 27.7% of the average scientist’s scientific messages to reading and knowledge acquisition. Anthropic’s study of 207,049 research sessions on consumer plans found that share at 7.6% of sessions on Claude and 3.7% on Claude Code, with data analysis much larger: 33.7% of Claude sessions and 48.1% of Claude Code sessions. The OpenAI paper tests two explanations. Weighting does not close the gap: pooling messages instead of averaging per scientist lowers the reading share only to 24.1%. Field mix does not either, since reading ranges from 23.8% to 32.8% across all seven fields. The authors suggest the difference may reflect researchers using ChatGPT more for knowledge acquisition and Claude more for computation, or differences in who is sampled: OpenAI’s group is verified academics, while Anthropic classified sessions as scientific without restricting to academic users. The units also differ: OpenAI counts messages, Anthropic counts sessions.

OpenAI, Science at the FrontierAnthropic, How scientists use Claude
Published16 November 2026 (paper dated 6 November)5 November 2026
Who is sampled1,819 verified academic scientists in a free-access programmeSessions on consumer plans classified as scientific work
Period20 August to 22 September 20262 August to 30 September 2026
Unit580,186 messages207,049 sessions
ClassificationGPT-5.6-Terra; 10 workflows, 40 tasks adapted from SciNetTwo classifiers; SciNet’s 7,262 tasks across 34 fields
Reading and knowledge acquisition27.7% of messages7.6% of Claude sessions, 3.7% of Claude Code
Survey251 programme participants528 scientists, mainly life and health sciences, active AI adopters

Ideation: small share, wide use

Ideation is where OpenAI’s summary and its paper put the most emphasis, and where the numbers most need their definitions. The web page says “68% of scientists used ChatGPT to develop research ideas.” The paper’s Table A3 shows what that means: 67.6% of scientists sent at least three ideation messages in the 34 days. At a threshold of one message, the figure is 84.1%. At ten messages it is 40.2%, and at 25 it is 19.9%. Ideation is only 5% of the average scientist’s messages. The paper reports that 42.7% of scientists averaged more than two ideation messages a week and 11% more than ten.

The paper splits ideation into “idea exploration” (identifying gaps in the literature, brainstorming directions, formulating questions from observations) and “question development” (generating hypotheses, formulating problems, models or approaches). Most ideation messages are in the second stage: formulating problems, models or approaches alone accounts for 73.8% of them.

The division of labour differs by stage. When exploring, scientists mostly ask the AI to propose what to look at: 85.1% of messages about identifying gaps in the literature and 69.3% of brainstorming messages were classified as “AI-formulated.” When developing a question, scientists mostly bring their own: 72.6% of hypothesis-generation messages were scientist-supplied. In formulating problems and models, 48.1% asked the AI to refine a supplied idea and 39.4% were scientist-supplied. OpenAI’s web page summarises the overall split as 35% led by the scientist, 36% developed together, 23% suggested by ChatGPT and 6% other.

The authors draw out a consequence that the summary page does not stress. Because scientists delegate part of the search for directions, “the set of ideas scientists are exposed to in the idea exploration stage is determined, in part, by AI.” They raise the risk that AI suggestions could steer many scientists toward the same problems, citing evidence associating AI use with “a narrower collective research agenda.” The data here cannot test that; it describes requests, not the ideas that resulted.

How often scientists push back

The paper’s “involvement score” measures how often scientists come back to an output. For each initial request, it checks whether the next scientific message in the same conversation, within seven days, asks for clarification, checking or revision of the output. Requests to extend the output (“now put this estimate in the draft”) and simple reactions do not count. Across workflows, the average scientist’s score is 31.8%.

WorkflowInvolvement score
Writing and communication38.0%
Scientific analysis34.4%
Research software and code33.7%
All workflows, average scientist31.8%

Scores from the paper’s Section 3.3.2, covering 125,947 initial requests from 1,809 scientists sent between 20 August and 15 September.

Two patterns are informative. Involvement is higher where AI capability has attracted most attention: writing, analysis and code. And within the same task, involvement is higher when the request needs more expertise: in 28 of 32 tasks, requests classified as needing advanced or frontier expertise drew more follow-ups, by 6.5 percentage points on average. Longer tasks drew only slightly more follow-ups, 2.6 points on average across 34 tasks. The paper reads this as evidence that “AI output fails to meet scientists’ standards more often in advanced research tasks.”

The authors interpret a follow-up as “evidence that ChatGPT did not execute the research task to the scientist’s desired standards.” That is a reasonable proxy with known gaps, and the paper names the main one: scientists may refine output outside the application, in their own editor, notebook or code. A scientist who silently discards a wrong answer, or who accepts a wrong one, scores zero. The score measures visible dissatisfaction within ChatGPT, not the correctness of what was used.

The survey points the same way. Asked what made it hard to get value from the programme, 45.4% of respondents chose “time needed to check or correct outputs,” the most common answer.

End-to-end autonomous research, in which a scientist asks the AI to run literature review, ideation and analysis as one job, is rare. About 15% of scientists sent at least one such request, the highest share in mathematics, but these requests are 0.4% of the average scientist’s messages, and the paper’s return-rate data suggests few came back to it. The paper connects this to the involvement result: with nearly a third of requests prompting correction, a fully automated project would let problems “go unaddressed and compound.”

The survey numbers

The survey ran from 30 September to 7 October 2026. Of 270 valid completed responses, 251 respondents consented to inclusion. The paper does not say how many participants were invited, so the response rate is unknown.

The discovery question asked respondents to imagine an important discovery that would take 10 years without AI and say how much sooner it would come with AI. Of 250 who answered, 97.6% expected some acceleration (98% on the web page), 91.2% at least one year sooner, 72.0% at least three years sooner and 21.6% seven to 10 years sooner. Mathematicians and statisticians were the most optimistic: 31% chose seven to 10 years, against 10% in engineering and materials science.

Discovery that would take 10 years arrives with AIAll fields (n=250)Mathematics and statistics (n=87)Life and health sciences (n=59)
Within 3 years22%31%20%
4 to 7 years50%41%44%
8 to 9 years19%16%29%
Just under 10 years6%10%0%
No acceleration2%1%7%

From the chart data on OpenAI’s study page. Respondents could choose more than one field.

Asked why, respondents mostly described speed. Of 367 claims extracted from 224 written answers, 52% concerned faster research execution, 23% better search for promising ideas and 14% pursuing a wider range of questions. Within the speed category, “faster task completion” alone was 49.6% of all claims. “Automating research loops” was 2.2%.

The productivity question is the source of the 1.9 times figure. For each workflow they had used AI for in the past 30 days, respondents estimated how much more or less time they would have needed without AI “to complete the same amount of work at the same quality.” Answers were options from “less time” to “400% or more.” The paper converts them to ratios: “less time” is coded as 0.5, “the same time” as 1, “400% or more” as 5, and respondents who did not use AI for a workflow receive 1 for it. Mean ratios ranged from 1.35 for administrative work to 2.83 for research software. OpenAI’s web page summarises the result as 1.9 times, a figure the paper itself does not report or derive; the key-findings PDF calls it “nearly twice as quickly” across eight activities. The paper also finds that workflows where respondents use AI for more of their time show larger reported gains, a correlation of 0.96 across eight workflow averages, and cautions that the correlation rests on only eight points.

Finally, the bottleneck question. Even with AI tools, 51.8% named limited time or staffing and 43.8% limited funding, with 71% choosing at least one of the two. Only 6.4% reported no major bottleneck.

What the headline numbers mean

The figures most likely to be quoted from this release each carry a definition that changes their meaning.

HeadlineWhere it appearsWhat it measuresWhat it does not measure
1.9 times as long without AIOpenAI web pageRespondents’ estimates of time without AI against time with AI, averaged across eight workflows (the paper reports means of 1.35 to 2.83 and codes non-use as 1, but does not show how 1.9 was computed)Measured time, output, or quality of work
98% expect faster discoveryWeb page and key findingsSurvey answer to a hypothetical about one 10-year discoveryAny observed change in discovery rates
68% used ChatGPT for ideationWeb page and key findingsScientists with at least 3 ideation messages in 34 daysWhether AI-generated ideas were pursued or useful
About 1 in 3 responses drew follow-upsKey findingsNext scientific message in the conversation sought clarification, checking or revisionErrors scientists did not flag, or corrections made outside ChatGPT
580,000 messages from over 1,800 scientistsAll three documentsClassified scientific messages from programme participantsScientists outside the programme or outside the seven fields

A worked example shows why a self-reported ratio cannot be read as a productivity measure, using invented numbers rather than the paper’s data. Suppose a researcher spends 10 hours a week coding with AI and believes the work would have taken 28 hours without it, close to the paper’s 2.83 times average for research software. If the AI made it practical to write analysis code the researcher would never have written by hand, the 28 hours were never going to be spent; the researcher is doing different work, not the same work faster. If the researcher then spends four hours checking and debugging the AI’s code, and recalls the session as “coding with AI,” that time may or may not be counted in the estimate. The paper itself notes that some interviewees described using ChatGPT to build research tools that would “otherwise have been impractical.” A ratio built on “the same amount of work at the same quality” is hard to answer for work that would not have existed.

What the study can and cannot tell us

The paper is unusually direct about its limits. In its productivity section it says respondents “are a selected group of scientists who applied for expanded AI access and chose to answer a survey about AI use,” and that “the findings may therefore not generalize to the broader population of scientists.” It says message shares “describe how scientists use AI, rather than the value AI provides or how scientists allocate their research time.” And it says that if scientists use AI tools other than ChatGPT, low shares in a workflow may simply reflect use elsewhere.

Read with those limits, the study supports several solid conclusions about this group. Its members use ChatGPT mostly to understand material and to write. Many of them use it to explore research directions, often letting it suggest what to look at. They frequently challenge and correct its output, more so for harder tasks. They believe it saves them time and will speed discovery, and they say time, staff and money still constrain them more than tools do.

It does not support claims about scientists in general, about the size of productivity gains, or about effects on discovery. There is no comparison group of similar scientists without access, no measure of research output such as papers, results or replication, and no follow-up beyond 34 days. The classifier’s accuracy is not reported against human labels. The survey’s response rate is not reported. And the study is produced by the company that runs the programme and sells the product, with all three authors listing OpenAI as an affiliation: one at OpenAI, one at Duke University and OpenAI, and one at the University of California, Berkeley and OpenAI, identified in the paper as a paid OpenAI contractor. That is not a reason to distrust the numbers, which are carefully defined, but it is a reason to want the same questions asked by an independent team with access to the same data.

The study also cannot settle the cross-company comparison it starts. The reading-share gap between OpenAI’s and Anthropic’s samples is large, and neither company’s data can tell whether it comes from the products, the users or the classifiers. Our survey of AI for biology’s evidence levels makes a related point about agentic science claims: describing what a system was used for is a different level of evidence from showing what it produced. And the involvement findings echo the argument in our piece on the verification bottleneck in AI review, that checking output becomes the binding cost as generation gets cheaper.

What the expanded programme offers

OpenAI’s 16 November post, “Broadening access to AI for scientists,” says the programme will give “up to 200,000 researchers worldwide free access to ChatGPT.” The next round’s applications open on Monday 23 November 2026. OpenAI will award 50,000 grants “on a first-come, first-served basis” to eligible researchers. Each grant covers the recipient and up to two collaborators on ChatGPT Plus, which OpenAI says supports up to 150,000 researchers. The post says 20,000 researchers in 33 countries have joined since the pilot began in July.

Summer pilotRound opening 23 November
SelectionLottery among eligible applicantsFirst 50,000 eligible applications
PlanChatGPT ProChatGPT Plus
Seats per grantApplicant plus up to four colleaguesApplicant plus up to two collaborators
DurationNot stated in sources readUp to 12 months (help page)
API creditsNot stated in sources readNot included (help page)
Data useNot used for training by defaultNot used for training by default

Sources: the paper’s Appendix A.1, OpenAI’s programme post and OpenAI’s help-centre page.

The switch from Pro to Plus is explained by the pilot’s data: “When we reviewed pilot usage, we found that most participants would have stayed within ChatGPT Plus usage limits.” OpenAI says it moved away from the lottery “based on feedback from researchers.” The help page adds terms the post does not: approved researchers receive “three complimentary Plus seats for up to 12 months”; collaborators may include research students at the same institution; collaborators from other institutions are not allowed by default; the offer “does not include OpenAI API credits”; and workspace admins can view, export, audit and delete workspace data and can opt in to sharing it to improve OpenAI’s models. Eligibility is unchanged in substance: research faculty or postdocs at eligible institutions, recent published research in one of the seven fields, and an intention to use ChatGPT for academic research. “Meeting the eligibility criteria does not guarantee approval,” the help page says.

Two arithmetic points follow from OpenAI’s numbers, and both are ours. First, 150,000 new seats plus the 20,000 researchers already enrolled comes to 170,000, short of the 200,000 goal, so reaching it implies further rounds or other routes. Second, a first-come process rewards applicants who are watching on the day applications open, which may shift the next cohort toward researchers already engaged with AI. That matters for any future study drawn from the programme, which would inherit a different selection from the pilot’s lottery.

The post also lists other routes. Universities can buy ChatGPT Edu for their campuses. For biology, GPT-Rosalind “brings together specialized biological reasoning and access to scientific tools and databases,” with access requiring approval through OpenAI’s trusted-access process; the Rosalind Academic Plan offers eligible labs at US universities “up to 30 seats at no cost for up to 12 months.” In mathematics, OpenAI says it contributed $5 million to a Simons Foundation initiative to help mathematicians examine and build on AI-produced results.

The science blog

The 15 November post, “Our new blog for science,” says the blog will publish three kinds of material: research results from OpenAI and collaborators, “methods and practical guidance” including “how to check the work and where a method falls short,” and first-hand perspectives from researchers. It summarises recent results OpenAI attributes to its models: a proof addressing the Navier–Stokes existence and smoothness problem in September, mathematical results in early October including one on what it calls the quasi-Riemann hypothesis, results later in October across mathematics, computer science and physics, and an analysis with PerturbAI and Scripps Research in which “thousands of agents powered by GPT‑Rosalind analyzed existing experimental data from millions of mouse brain cells in 32 minutes.”

Those are OpenAI’s descriptions of its own work. Some have been contested. The South China Morning Post reported on 16 November that OpenAI’s October claims of discoveries in mathematics, computer science and physics “drew pushback from the scientific community, particularly mathematicians.” The blog post itself acknowledges open questions, noting that systems that generate proofs “raise questions about how new arguments are checked, explained, and incorporated into the field.” A blog that publishes failures and checking methods, as promised, would be a useful complement to the usage study; whether it does is something readers can track.

What would strengthen the evidence

Several additions, none of which OpenAI has announced, would let the study’s findings carry more weight. Each is a proposal from The Frontier Wire.

A classifier audit, comparing GPT-5.6-Terra’s labels with expert labels on a consented or synthetic sample, would show how reliable the task and ideation categories are.

A comparison group would address selection. The waitlist from the pilot’s lottery is a natural one: eligible applicants who lost the lottery are similar to those who won, and their publications, grant applications or self-reports over the same period could be compared.

Output measures would move the study from use to effect: papers, preprints, code releases or grant submissions over a longer window, for participants and the comparison group.

Independent access would address the conflict of interest. The paper’s privacy design, in which analysts see only labels, could in principle be extended to outside researchers working under the same rules.

Until then, the defensible reading is the paper’s own. Scientists in this programme use ChatGPT heavily for reading and writing, increasingly for exploring ideas, and with a good deal of checking. They expect it to speed discovery. Whether it does is a question the data published so far was not designed to answer.

Sources

  1. OpenAI: Science at the Frontier, How scientists use ChatGPT (16 November 2026)
  2. Liao, Shan and Chatterji: Science at the Frontier, How Scientists Use ChatGPT (paper, dated 6 November 2026)
  3. OpenAI: Science at the Frontier, key findings overview (PDF)
  4. OpenAI: Broadening access to AI for scientists (16 November 2026)
  5. OpenAI: Our new blog for science (15 November 2026)
  6. OpenAI Help Center: ChatGPT for Academic Researchers
  7. OpenAI: AI for scientific research and discovery
  8. Anthropic: How scientists use Claude (5 November 2026)
  9. South China Morning Post: China's Moonshot AI puts Kimi models into race for scientific discoveries (16 November 2026)

Questions readers ask

What data is the Science at the Frontier study based on?

Two datasets. The first is 580,186 messages classified as scientific, sent between 20 August and 22 September 2026 by 1,819 scientists on the ChatGPT for Academic Researchers plan, through ChatGPT and Codex. The second is a survey run from 30 September to 7 October 2026, with 251 consenting respondents.

Did OpenAI's researchers read scientists' messages?

No, according to the paper. A model, GPT-5.6-Terra, classified each message into predefined categories, and the researchers analysed only those labels. The paper says no researcher observed or had access to the underlying messages.

Does the study show that AI doubles scientists' productivity?

No. The 1.9 times figure on OpenAI's web page comes from survey answers in which respondents estimated how long the same work would have taken without AI. It is a self-reported estimate from a self-selected group, not a measured change in output.

What does the expanded ChatGPT for Academic Researchers programme offer?

OpenAI says it will award 50,000 grants on a first-come, first-served basis when applications open on 23 November 2026. Each grant gives the recipient and up to two collaborators free ChatGPT Plus; OpenAI's help page says the seats last up to 12 months and do not include API credits.

Did The Frontier Wire verify the findings independently?

No. The underlying data is not public. This article reads OpenAI's full 76-page paper, its key-findings summary, the three OpenAI posts, the programme's help page and Anthropic's comparable study, and labels any arithmetic as its own.

More research