Roughly 1.3 million people already use ChatGPT every week for advanced science and mathematics. They send about 8.4 million messages doing it.
OpenAI published both figures on Wednesday, buried in the middle of a blog post announcing that it will hand 100,000 of those people its frontier models at no charge. The numbers do more work than the announcement does. They say the adoption already happened, and the program is OpenAI formalizing a behavior it can measure.
The offer itself is straightforward. ChatGPT for Academic Researchers opens with 10,000 seats this summer, at institutions that already include the Institute for Advanced Study and France's École normale supérieure, and expands to 100,000 researchers through 2027. Approved researchers get frontier model access across ChatGPT, ChatGPT Work and Codex, and each can invite up to four collaborators from the same institution. Axios, which got the announcement first, put the value at roughly a year of access equivalent to a $200-per-month ChatGPT Pro account.
Then there is the part OpenAI did not put in the headline. The program is aimed at biology, chemistry, physics, mathematics and engineering. It is not aimed at the people studying the models.
What a Seat Actually Includes
The package is broader than a Pro subscription with a coupon on it.
| Component | What participants get |
|---|---|
| Models | The GPT-5.6 family at launch, including Sol Pro |
| Products | ChatGPT, ChatGPT Work, and Codex |
| Limits | Expanded deep research, higher usage limits, larger context windows |
| Domain tooling | More than 75 life science skills across genetics, genomics, sequencing, single-cell analysis, protein modeling and drug discovery |
| Connectors | Scientific literature, public genomic and clinical databases, satellite imagery, computational notebooks, data platforms, reference managers |
| Privacy | Business-grade protections; data not used to train models by default |
| Collaborators | Up to four per approved researcher, each counting toward the program total |
OpenAI splits the GPT-5.6 line by cost and difficulty rather than by capability tier alone. Terra balances capability against efficiency for everyday work, Luna answers faster for lighter tasks, and Sol handles the hardest scientific and mathematical problems. On FrontierMath Tier 4, which tests research-level mathematical reasoning, OpenAI reports GPT-5.6 Sol at 83% against 72.5% for GPT-5.5. On GeneBench Pro, a test of complex biological data analysis, the company says GPT-5.6 Sol Pro solves 31.5% of tasks.
That second number is the honest one. A third of a biology benchmark is not a research assistant. It is a tool with a known failure rate that a scientist has to check, which is roughly what OpenAI's own evaluation work has been saying since it published the first real test of AI scientists and watched its best model pass 36 percent.
Eligibility runs through the institution. Applicants must be at recognized, degree-granting colleges or universities with a high level of research activity, verify their affiliation, and describe their active research and intended scientific use. Institutions already running ChatGPT Edu will have the free access coordinated through their existing workspace.
The Usage Data Is More Interesting Than the Giveaway
OpenAI released a slice of its own telemetry alongside the announcement, covering scientific requests from June 21 to July 20, 2026. One finding is worth pulling out of the charts.
Researchers in the top 20% of AI usage within their own field are almost twice as likely as their peers to hand the model a task the company estimates would take a human four hours or more. Nearly 7% of their requests fall in that bucket, against 3.5% for other researchers in the same field.
That is a behavioral gradient, not a capability one. Same models, same field, different expectations about what to delegate. The heavy users are not getting a better tool. They have recalibrated what counts as a reasonable ask, and the gap between the two groups is the entire practical story of AI adoption in research right now.
OpenAI also claims the shift is showing up in the literature, with a growing number of papers acknowledging ChatGPT's contribution, and flags its own caveat: the July figure is incomplete because the full-text index it used extends only through July 21.
Two named examples appear in the post, both presented by OpenAI rather than independently verified. Physicist Rogerio Jorge and his team are using AI to develop open-source fusion research software used by industry and national laboratories to design fusion energy devices. In theoretical computer science, Barna Saha, Yinzhan Xu and Christopher Ye used GPT-5.5 Pro to develop a proof establishing new limits on how efficiently computers can solve high-dimensional geometry problems, then validated and refined the results themselves.
That last clause is the one that matters, and OpenAI wrote it in. The humans checked the work. When OpenAI skipped that step on a mathematics claim earlier this year, the correction was public and expensive, a pattern we traced in OpenAI's last math claim was an embarrassment.
The Researchers Who Study AI Were Left Out on Purpose
The clearest criticism of the program came from Axios in the same story that broke it.
The program targets math and the natural sciences. It does not address what AI researchers themselves have been asking for, which is access to model weights and the training data underneath them. Inference access, however generous, does not enable the work that field needs to do.
Both sides of that argument are real:
- OpenAI and Anthropic's position is that restricting access to weights limits misuse. A published set of frontier weights cannot be recalled, and the past year has produced enough incidents involving autonomous agents to make that argument concrete rather than hypothetical.
- The researchers' position is that limited access hampers independent evaluation of model behavior, reproducibility of published results, and safety work. You cannot audit what you can only query.
Axios summarized the tension in one line: OpenAI is expanding researchers' access to its most capable tools while keeping the underlying systems largely closed.
The timing sharpens it. Twenty-plus companies including Nvidia, Meta, Microsoft, Google and OpenAI signed an industry letter this month arguing that open weights serve American AI leadership, a fight we covered when Anthropic declined to sign and published its own answer instead. Signing a letter in favor of open weights and running a 100,000-seat program that ships none is not a contradiction, exactly. It is a distinction between what a company advocates and what it distributes.
What This Changes for Academic ML Work
For a data scientist or ML engineer inside a university, the practical effects are narrow and specific.
- Compute-poor labs get frontier inference, not frontier training. A free Pro-tier account for a year removes a real budget constraint on analysis, literature work and code. It does nothing for anyone who needs to train or fine-tune.
- The collaborator multiplier is the sleeper feature. Four invitees per approved researcher means a single approval can cover a small group, and each invitee counts against the program cap. Early applicants get more of the pool.
- Codex access matters more than the chat interface. The published usage breakdown puts Codex on research execution and formal analysis: writing code, running it, analyzing results. That is where most of a computational scientist's day goes.
- Reproducibility does not improve. A paper whose analysis depended on a hosted model at a specific point in time is not reproducible in the sense the field means, and free access at scale increases the number of such papers.
- Interpretability and evaluation work is unaffected. If your research is about the model rather than done with the model, this program is not for you, and OpenAI has been explicit about that.
The Bottom Line
Strip out the press release and two things happened on Wednesday.
OpenAI removed a cost barrier for 100,000 scientists who, by its own telemetry, were already using the product 8.4 million times a week. That is a real transfer of capability to people with small grants and no enterprise contract, and calling it purely a distribution play understates it. It is also the cheapest possible version of that transfer, since the marginal cost of inference to OpenAI is not the sticker price of a Pro account.
What it does not do is change the terms on which the field can study these systems. The one request AI researchers have been making for three years, access to weights and training data, went unaddressed in a program explicitly framed around accelerating science. The company's rationale for that is defensible. The gap it leaves is also real, and no amount of free inference fills it.
The most quotable line in OpenAI's post was a statement of strategy: put capable tools in the hands of the research community and let researchers pursue the questions they know best. That works cleanly until the question a researcher knows best is what the model is doing.
Sources
- Accelerating scientific discovery with ChatGPT for Academic Researchers (OpenAI, Jul 29, 2026)
- Exclusive: OpenAI offers 100,000 academics free ChatGPT access (Axios, Jul 29, 2026)
- OpenAI opens new ChatGPT for Academic Researchers program to 100,000 scientists (SiliconANGLE, Jul 29, 2026)
- OpenAI will provide free AI models to select researchers (Engadget, Jul 29, 2026)
- OpenAI Launches Free AI Access for Scientists: Apply Now, Model Weights Still Off-Limits (Tech Times, Jul 29, 2026)
- OpenAI Launches Free ChatGPT Research Program For 100,000 Scientists (Dataconomy, Jul 30, 2026)
- OpenAI to provide free ChatGPT access to researchers: What's in the plan? (Business Standard, Jul 30, 2026)
- Introducing NextGenAI (OpenAI)
- Advancing the next era of national science (OpenAI)
- Life science research skills for ChatGPT (OpenAI, GitHub)