Anthropic is opening 10,000 seats on a new Claude tier for research staff. Scientists anywhere in the world get a year of free or discounted access: standard seats cost nothing, premium seats with five times the usage limits cost $15 a month. A lab that runs through its allowance can apply to the company's AI for Science program for up to $50,000 in compute credits on a single project. Anthropic says it plans to raise the number of seats substantially over the coming months.
Access runs through a verification form. The applicant has to be a principal investigator or hold an equivalent position at an academic or nonprofit research organization; once confirmed, they can add the rest of their lab to the plan. Applications to AI for Science are open to any researcher.
This is a step in a sequence rather than a standalone launch. In June the company shipped Claude for Science, a product that bundles the tools researchers ask for most, produces verifiable results and gives flexible access to compute. AI for Science, which hands free compute credits to researchers working on significant projects, has run alongside it and has so far been weighted toward biology. That weighting now lifts: credits are open to other fields, including projects with heavy compute bills. The examples Anthropic offers are work that advanced the study of the Riemann zeta function and Claude's work in protein design.
The restrictions are more revealing than the seat count. For now, researchers in biology and chemistry will be able to use only Opus-class models. Claude Fable models will continue to block professional queries about biology and drug development, on dual-use grounds. For the work those refusals cover, Anthropic is building a separate route: a program with the US government that would give life-science specialists access to Mythos-class models for research and development. The first participants are already connected, and the company says it will describe the program in more detail and widen access.
In the same announcement, Anthropic is releasing a research version of the Model Hardware Standard, a shared specification meant to let AI agents control physical devices safely. A first group of scientific labs and advanced manufacturers gets access. Who else writes that specification, and what it obliges a device to do, is not spelled out.
Two smaller items round out the post: a $5 million funding program for independent research into how AI affects user wellbeing, and a set of answers about the company's watermarking method — how it works, whether it affects Claude's outputs, and why it was introduced.
Ten thousand seats is the number in the headline and the least interesting figure in the post. Spread across every discipline and every country, it is a pilot, and Anthropic says as much by promising more. What the announcement actually describes is a graded access system: Opus for credentialed biologists and chemists, Fable refusing the same questions from everyone else, Mythos reachable only through a program run jointly with a government. That is a credentialing layer for frontier capability, and it is being assembled in public under the heading of a science program. I think the sequencing is defensible — gate first, widen later — but it means a company and a state are deciding, application by application, which researchers get to ask which questions.
Notably absent from the announcement is what happens in month thirteen. The free and discounted seats run for one year. A free year is also a year in which a lab's scripts, workflows and graduate students get built around one vendor, and the renewal price is the one number the post does not contain.
A biologist is at once the researcher Anthropic most wants to hand compute to and the one it most wants to restrain. The seats, the credits and the grants all point toward acceleration; the model gating and the government program point the other way, at the same desk.