Harvey, a $15 Billion AI Legal Startup, Lost Money Every Time a Lawyer Used It. A Chinese Model Saved It.

Harvey, the legal AI company last valued around $15.6 billion, spent part of this year running a business that got worse the more customers loved it. Bloomberg reported on September 21 that its gross margin fell from about 50% at the start of the year to -50% by June, according to a person familiar with the matter, after a March update to its AI agents made customer usage spike. In plain terms, Harvey was spending about a dollar and a half to deliver every dollar of software it sold. That figure has been passed around all week, usually attached to a tidy moral: Harvey built on OpenAI and Anthropic, the models cost too much, so Harvey built its own. The number is right. The moral is mostly wrong. Bloomberg

What actually went wrong, and it wasn't the models

Harvey sells law firms a seat. Fixed annual price, use it as much as you want. Its model providers charge by the token. Those two facts can live together peacefully until usage jumps, and in March, Harvey shipped an agent update that made usage jump hard. As reporting compiled by The Daily Brief lays out, Harvey has seen a twentyfold rise in token usage this year while annual recurring revenue roughly doubled, from $190 million at the start of 2026 to more than $400 million. A partner running one query a week and an associate leaving a diligence agent grinding through a data room overnight now paid Harvey the same flat rate, while Harvey's metered bill from OpenAI and Anthropic climbed with every token. No flat seat price survives a cost curve that steepens twentyfold. The margin did not collapse because Harvey used frontier models. It collapsed because agent usage overran a pricing structure built for humans clicking buttons. THE D[AI]LY BRIEF

The fix came from Beijing

In August, Harvey released a model called Tenet, and this is the part that earns the double takes. Tenet is not built from scratch. Harvey's own engineering blog describes it as a Kimi K3 base that it post-trained together with Fireworks research for long-horizon legal work, Kimi K3 being a Chinese open-weight model from Moonshot AI. So an American legal startup that counts OpenAI among its investors now ships its flagship model on weights released out of Beijing. The South China Morning Post flagged the same pivot: a company that previously focused on customising closed proprietary models from Anthropic, OpenAI and Google moved its headline product onto open weights. HarveySouth China Morning Post

The economics explain everything. An open-weight model is one you can download and run on your own hardware, paying for compute instead of renting intelligence by the token. Harvey's blog says the post-trained model delivers better answers on some review work at roughly one-tenth the cost per cell and cut cost per query on one feature by about 90%, and Bloomberg reports margins turned positive again after the switch.

Harvey did not actually leave the frontier labs

Here is the caveat the victory laps skip. Neither Bloomberg nor anyone else has published Harvey's margin after Tenet, or the share of its work that still routes to GPT and Claude. As daily.dev's summary of the reporting notes, Anthropic has flagged the trend internally while noting Harvey still needs Opus for hard tasks, and Harvey's platform lets firms route across models and choose which ones are enabled. Tenet took the high-volume, repetitive work off the expensive meter. The genuinely hard reasoning appears to have stayed where it was. That is not independence from the frontier labs. It is triage, and the distinction matters if you are trying to copy the playbook. daily.dev

Harvey has a lot of company

The reason this rattled people is that Harvey is not alone. In the same Bloomberg reporting and the coverage around it, startups spanning legal, healthcare, financial and customer service, including Abridge, Decagon and Ramp, have announced plans to develop or customize their own models, with some firms already routing 80% of their traffic to in-house models, and investors such as Sequoia Capital and General Catalyst backing the trend. Both OpenAI and Anthropic are moving toward IPOs, and their most successful application-layer customers are now also their most motivated to leave. Every company that priced its product against frontier tokens just watched Harvey's negative 50% and reached for a calculator. All-Weather Media

The skeptics think this is theater

Not everyone buys the revolution. As daily.dev reported, a Menlo Ventures partner called the emphasis on owning a model largely "cosplay", and the point lands. Post-training an open-weight base is not the same as building frontier intelligence, and most companies will land on a hybrid: a cheap in-house model for the bulk of the work, rented frontier models for the peaks. Owning the easy 80% of your traffic is not sovereignty over your stack, and treating it that way is how founders talk themselves into a hard research problem they did not need to take on. daily.dev

The lesson everyone is drawing, and the one that actually holds

The tidy version says the takeaway is build your own model. The version that survives the fact-check is quieter and more useful. Model routing stops being an optimization and becomes the core of the business, because deciding which query deserves a frontier token and which does not is now a margin decision made thousands of times an hour. Pricing can no longer be set once and left alone, because the product underneath it consumes cost on its own schedule now, not the customer's. And a healthy gross margin in AI means very little until you know how hard people are actually using the thing.

Harvey's real story is not that it built a model. It is that in agentic software the meter never stops running, and any company still selling flat rates on top of it is one usage spike away from its own June.