Harvey, the legal AI startup valued at $15.6 billion, has moved its flagship product off frontier models from OpenAI and Anthropic and onto Tenet, an in-house system post-trained on Moonshot AI’s open-weight Kimi K3, Bloomberg reported on 21 September. The switch followed a period in which Harvey’s gross margins collapsed from roughly 50 percent at the start of the year to negative 50 percent by June, even as customer usage climbed sharply after a March update to its AI agents.
The economics were straightforward and brutal. More usage on proprietary frontier models meant more inference cost passed straight through, and Harvey was absorbing the difference rather than passing all of it to law firm customers. Margins turned positive again only after Tenet launched in August, built on a Chinese open-weight model that Harvey could run and customise on its own infrastructure rather than pay per token to a lab it does not control.
Harvey is not alone. Bloomberg names Abridge, Decagon and Ramp, backed by investors including Sequoia Capital and General Catalyst, as pursuing similar pivots toward open-weight models, several originating from Chinese labs, as a way of decoupling margins from the pricing decisions of OpenAI and Anthropic. It marks a shift from treating frontier labs as infrastructure partners to treating them as a cost centre to be engineered around wherever model quality allows it.
What the shift does not resolve is how much of Harvey’s product quality depended on the frontier model it just walked away from for its highest-value work, and how customers judging legal AI on accuracy rather than cost will view the swap. Bloomberg’s reporting covers the margin recovery in detail. It says less about whether output quality moved with it.
