The claim
A general model becomes useful in science at the point where its output can be checked: by a formal proof system, by a wet lab (a partner's, or now its own), or through a shared instrument interface.
What has to hold for this route to work.
The real bottleneck. The bottleneck is placed on verification rather than generation: producing a candidate is not the hard part, giving anyone a reason to trust it is. Almost every Anthropic science artifact in 2026 points the same way: make the claim checkable. Fermat's Last Theorem was made machine-verifiable, the Riemann-related bound shipped with a formal proof, and MHS turns rebuilding every agent-instrument integration into a shared interface.
What is established. Long-horizon autonomous formal work is established, and model-assisted design has been validated by external wet labs. Eleven days, 13 million lines and 29,500 intermediate theorems is the longest publicly documented autonomous formalization run; binders for 14 of 15 targets were validated by Adaptyv Bio and Twist Bioscience, putting the result outside the company.
What is still a claim. That verification-first generalizes is still a claim: the interface has named vendors but remains a research preview, and the brightest mathematical result came from a model outsiders cannot run. MHS has named vendors connected (AWS, Danaher, QIAGEN, Tecan, Universal Robots) and test sites at Genentech, Carnegie Mellon and QuEra, where CMU cut a setup that normally takes weeks to about eight hours; it is still a research preview, and vendors building support is not the same as routine use. The Riemann-related bound (41.6% to 67.2%) came from an unreleased research model that outsiders cannot reproduce: they can only check the proof it emitted.
The chosen product abstraction. The product abstraction is a general model plus a checking layer: no domain models, no wet lab of its own: distribute access and an interface standard instead. Discounted access turns a frontier model into research infrastructure, and raises the same allocation questions as compute, data and instruments; MHS aims to be the public layer between agents and instruments. Neither is a science product; both are what other people's science products would sit on.
