What you actually get for free
Start with the contents. This is not simply a batch of free accounts.
Models and products in scope
Participants receive free access to frontier models across ChatGPT, ChatGPT Work, and Codex, including the GPT-5.6 family at launch.
Three extensions come with it: expanded deep research (the feature where the AI investigates multiple sources on its own and returns a consolidated report), higher usage limits, and larger context windows (the cap on how much text can be read at once). OpenAI describes the combination as supporting scientific reasoning, agentic execution, and research workflows across disciplines.
The announcement also spells out which model to reach for. GPT-5.6 Terra balances capability and efficiency for everyday research, GPT-5.6 Luna answers lighter-weight tasks faster, and GPT-5.6 Sol takes on the hardest scientific and mathematical problems. We cover the pricing and positioning of the three in a separate article.
Participants will receive free access to our frontier models across ChatGPT, ChatGPT Work, and Codex, including the GPT-5.6 family of models at launch. They will also have expanded deep research, higher usage limits, and larger context windows. Together, these capabilities support scientific reasoning, agentic execution, and research workflows across disciplines. / GPT-5.6 Terra balances capability and efficiency for everyday research, GPT-5.6 Luna provides faster responses for lighter-weight tasks, and GPT-5.6 Sol tackles the most difficult scientific and mathematical problems. — From the statements on what is included and on how the three models divide the work
The benchmark numbers cited
Two figures are offered as evidence. On FrontierMath Tier 4, which measures research-level mathematical reasoning, GPT-5.6 Sol scores 83%, against 72.5% for GPT-5.5.
On GeneBench Pro, which evaluates complex biological data analysis and scientific reasoning, GPT-5.6 Sol Pro solves 31.5% of tasks. Set against the mathematics figure, that second number also shows how much room is left in biological data analysis.
On FrontierMath Tier 4, which measures research-level mathematical reasoning, GPT-5.6 Sol scores 83%, compared with 72.5% for GPT-5.5. On GeneBench Pro, which evaluates complex biological data analysis and scientific reasoning, GPT-5.6 Sol Pro solves 31.5% of tasks. — From the statement on the two benchmark results
Domain skills and external connectors
This goes past general-purpose chat. More than 75 life science skills are available, spanning genetics, genomics, sequencing, single-cell analysis, protein modeling, and drug discovery.
Connectors are included as well, providing access to scientific literature, public genomic and clinical databases, satellite imagery, computational notebooks, data platforms, and reference managers.
The division of labor is stated explicitly. Codex helps write and debug code, analyze datasets, and build reproducible workflows, while ChatGPT Work supports longer projects: finding funding opportunities, preparing grant applications, reviewing literature, drafting manuscripts, and creating materials to communicate results.
Training is part of the package too. The program will offer training tailored to different levels of experience, from getting started and improving workflows through to advanced research applications.
Papers and grant paperwork usually arrive as PDFs, and converting them into something easier to handle cuts the friction of feeding them to a model.
Researchers can use more than 75 life science skills spanning genetics, genomics, sequencing, single-cell analysis, protein modeling, and drug discovery. Connectors support research across disciplines, providing access to scientific literature, public genomic and clinical databases, satellite imagery, computational notebooks, data platforms, and reference managers. / Codex can help write and debug code, analyze datasets, and build reproducible workflows. ChatGPT Work can support longer projects such as finding funding opportunities, preparing grant applications, reviewing literature, drafting manuscripts, and creating materials to communicate results. / The program will offer training tailored to different levels of experience, from getting started and improving workflows to advanced research applications. — From the statements on available skills and connectors, on how ChatGPT Work and Codex divide the work, and on training
Who qualifies, and how to apply
The conditions are written plainly. This is not a program any individual can sign up for.
Degree-granting institutions with "a high level of research activity"
The initial program is open to qualifying researchers at selected academic institutions. Eligible institutions must be recognized, degree-granting colleges or universities with a high level of research activity.
The fields named are the sciences, mathematics, and engineering, and the intended uses run broadly, from preparing grant applications to testing hypotheses.
The program will help researchers across the sciences, mathematics, and engineering take on advanced problems, accelerate discovery, and improve productivity, from preparing grant applications to testing hypotheses. — From the statement on the fields covered and the range of intended uses
What the application requires
Applicants need two things: verification of their institutional affiliation, and information about their active research and intended scientific use.
Approved researchers may then invite up to four collaborators from their institution. There is a catch: each collaborator must also verify their affiliation, and each one counts toward the program's total number of accounts. One approval effectively claims five slots.
For institutions that already have ChatGPT Edu, free access granted through this program will be coordinated through the institution's workspace. Applications opened the same day as the announcement.
The initial program is open to qualifying researchers at selected academic institutions. Eligible institutions must be recognized, degree-granting colleges or universities with a high level of research activity. / Applicants will need to verify their institutional affiliation and provide information about their active research and intended scientific use. Approved researchers may invite up to four collaborators from their institution. Each collaborator must also verify their affiliation and counts toward the program's total number of accounts. / For institutions with ChatGPT Edu, free access granted through this program will be coordinated through the institution's workspace. / Applications are open today. — From the statements on eligible institutions and researchers, on what applying requires, and on institutions with ChatGPT Edu
Scale and schedule
The rollout is staged. It starts with 10,000 researchers this summer, with access already available at institutions such as the Institute for Advanced Study (IAS) and École normale supérieure (ENS). From there the plan is to expand to 100,000 researchers through 2027.
Data handling is spelled out as well. Workspaces include business-grade privacy and security protections, and data is not used to train the models by default. When research data is involved, that is a condition worth reading before anything else.
We're introducing ChatGPT for Academic Researchers, a program that will give 100,000 researchers at selected academic institutions free access to our frontier models. / We're starting with 10,000 researchers this summer, with access already available at institutions such as the Institute for Advanced Study (IAS) and École normale supérieure (ENS). We plan to expand to 100,000 researchers through 2027. / Workspaces include business-grade privacy and security protections, and data is not used to train our models by default. — From the statements on the program's scale, its start and expansion plan, and data handling
Why give this much away now
The announcement backs the timing with usage data from research settings.
1.3 million people a week already use it for science and math
The published figures: roughly 1.3 million people use ChatGPT for advanced science and mathematics each week, generating about 8.4 million messages.
The shift is described as most visible in mathematics. In the past six months, AI moved from occasional use on isolated problems to a more regular part of mathematical research. A growing number of papers now acknowledge ChatGPT's contribution.
Heavier users hand over heavier work
One more figure stands out. Researchers in the top 20% of AI usage within their field are almost twice as likely as their peers to ask AI to take on tasks estimated to require four hours or more.
Concretely, that is nearly 7% of their requests, against 3.5% among other researchers in the same field. The more people use these tools, the larger the unit of work they delegate — an observation that travels well beyond any single field.
That results depend not only on raw capability but on how the tool is set up echoes OpenAI's own case where harness settings tripled a score, the harness being the scaffolding a model runs inside.
AI is becoming a more capable research tool faster than many expected. Each week, roughly 1.3 million people use ChatGPT for advanced science and mathematics, generating about 8.4 million messages. / The shift is especially visible in mathematics. In the past six months, AI has moved from occasional use on isolated problems to a more regular part of mathematical research. A growing number of papers now acknowledge ChatGPT's contribution, reflecting how quickly researchers are adopting these tools in their work. / Researchers using these tools most intensively are also taking on more ambitious work. Those in the top 20% of AI usage within their field are almost twice as likely as their peers to ask AI to take on tasks estimated to require four hours or more: nearly 7% of their requests, compared with 3.5% among other researchers in the same field. — From the statements on usage scale in research, the shift in mathematics, and the link between intensity of use and the weight of delegated work
One piece of a $250 million commitment
The program does not stand alone. ChatGPT for Academic Researchers is part of a commitment of more than $250 million through 2027 to support external scientific research and discovery.
That includes NextGenAI, a $50 million initiative supporting research institutions, and work with the Department of Energy's Genesis Mission to bring frontier AI to researchers at national laboratories and universities.
OpenAI also states its own position: the strategy is not to decide which scientific problems deserve attention or to try to solve them all itself, but to put capable tools in the hands of the research community and let researchers pursue the questions they know best.
ChatGPT for Academic Researchers is part of a commitment of more than $250 million through 2027 to support external scientific research and discovery. That includes NextGenAI, our $50 million initiative supporting research institutions, and our work with the Department of Energy's Genesis Mission to bring frontier AI to researchers at national laboratories and universities. / Our strategy is not to decide which scientific problems deserve attention or try to solve them all ourselves. It is to put capable tools in the hands of the research community and let researchers pursue the questions they know best. — From the statements on the scale and composition of the commitment, and on OpenAI's own strategy
Wrapping up: read the conditions before the headline number
The figure that matters here is not "100,000." It is that eligibility is gated at the institution level, and that one approval occupies up to five account slots. If your institution is not on the list, individual track record does not get you to the door.
The other thing worth checking is data handling. That data is not used for training by default is a precondition for anyone working with unpublished research. Read the other way, its absence in some other free offer is a reason to be careful.
As for the contents, the scaffolding around the models is thicker than the model access itself: more than 75 life science skills, connectors into literature and genomic databases, training matched to experience level. As with how context engineering is approached, the design principle showing through is that handing over a bare model is not what produces results — shaping what surrounds it to fit the work is.
