Harvey's gross margin hit minus 50% — so it stopped renting frontier AI
Two arguments about frontier AI landed the same afternoon: customers finally showing what the invoices do to a business, and a lab building a panel it says can challenge its research — just not slow it down.
Harvey's gross margin fell from about 50% at the start of this year to minus 50% by June, and only turned positive again after it released Tenet, its own model built on Moonshot's Kimi K3, according to a Bloomberg report. The legal AI company's token usage rose twentyfold over the same period, and because both OpenAI and Anthropic now bill enterprises for model usage on top of base subscription fees, that growth landed straight on the company's cost line. Harvey is the loudest case rather than the only one: Abridge, Decagon, Ramp and Rogo are reported to be making similar moves toward open-weight or in-house models, and Sequoia Capital and General Catalyst are funding the shift.
That is the open-weights argument in its most persuasive form — not a policy position, just an arithmetic one. Nobody in that cohort has to be ideologically committed to buying less from the frontier labs; they only have to look at a rising cost of goods sold and notice that the capability they need is available elsewhere at a fraction of the price. We covered Harvey's first move in August — Harvey launches Tenet, its first in-house legal model — and the same trade shows up in legal publishing, where Thomson Reuters builds its own legal AI instead of renting from OpenAI. The pattern is not disloyalty. It is the difference between a subscription and a fifth of your revenue.
OpenAI has set up an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, to assess and coordinate the release of results from the internal model that the company says has now resolved more than 100 long-standing open problems across most areas of mathematics. Nine mathematicians are named as initial members, among them Timothy Gowers, Martin Hairer, Edward Witten, Ravi Vakil and Melanie Matchett Wood. The group is unpaid, may publish advice OpenAI did not ask for, and can change its own membership.
The limits are written into the announcement. The group "will not be responsible for advising us on how to pace our internal progress on mathematics," and only one member, Camillo De Lellis, also signed the open letter from 25 Fields Medalists that warned the field is severely misaligned with AI companies. So the panel gets real authority over how results are reviewed and communicated, and none over how fast they arrive — which is precisely the lever the letter asked for. We wrote about the Navier–Stokes claim when it landed (OpenAI's AI solved Navier-Stokes — and a mathematician cries foul), and the gap between reviewing results and pacing them is the story now: a lab can buy credibility without buying a brake.
What to watch: whether the group's first public output criticises OpenAI's publication process or simply certifies it, and whether any frontier lab follows with a reviewer panel of its own.
If a lab's own panel can question its results but not its speed, does that count as oversight? Tell us in the comments.
