Harvey's margins hit minus 50%, so it moved onto Chinese open weights
Bloomberg reported on 21st Sep'26 that legal AI company Harvey watched its gross margins fall from roughly 50% to minus 50% by June, as token consumption rose twentyfold under OpenAI and Anthropic's usage-based enterprise pricing. Margins turned positive again in August after Harvey shipped an in-house model, Tenet, post-trained on Moonshot's open-weight Kimi K3 and built with Fireworks AI. Bloomberg names Abridge, Decagon and Ramp as pursuing the same open-weight pivot.
What it actually means
The number to sit with is minus 50%, and it was not caused by a price rise but by success. Harvey's customers used the product more and agentic workflows re-read the same context on every call, so consumption grew twentyfold while the contract did not. If you sell a flat subscription on top of metered tokens, your best customers are the ones losing you the most money, and that arithmetic does not improve as the product gets better.
The copyable part is the escape route, because Harvey did not renegotiate; it post-trained an open-weight base and moved its flagship onto it, which took a Chinese model and a serving partner rather than a research team of its own. That option is now open to anyone with a narrow domain and real usage data, and the frontier labs have lost the quiet assumption that their downstream applications have nowhere to go.
Your best customers are the ones losing you the most money, and that does not improve as the product gets better.
-50%Harvey's gross margin before it left the frontier APIs
Source: Bloomberg · Tuesday 22 September 2026