AI
Harvey Turned Its Margins Around by Building Its Own Model, and the Whole App Layer Is Watching
Here is a number that reshaped a company in six months. Harvey, the legal AI startup valued at about 15.6 billion dollars, watched its gross margins swing from about 50 percent at the start of the year to minus 50 percent by June. Token usage jumped twentyfold after a March update to its AI agent tools, and usage based pricing from OpenAI and Anthropic turned surging demand into a cost engine.
Harvey answered by building its own model. In August it released Tenet, customized from Moonshot's open weight Kimi K3 with Fireworks AI as the research partner, and margins turned positive again. Kimi K3 is a serious base, with 2.8 trillion parameters, 104 billion active per token, and a one million token context window, released by the Chinese lab Moonshot in July with open weights.
The numbers Harvey published are striking. Tenet completes nearly twice as many held out tasks on LAB and 20 percent more on LAB contracts than the base Kimi K3, lifting the all pass rate by 9 and 2 percentage points, and on one extraction task it improved answer quality by 3.6 points and citation quality by 12.1 points at roughly one tenth the cost per cell. The motive is explicit. Open weights mean cheaper per token prices.
Harvey is far from alone. Bloomberg reports that Abridge, Decagon, and Ramp are pursuing similar open weight pivots, with Sequoia Capital and General Catalyst backing the trend. The pattern is clear. Startups that grew up on frontier lab APIs are now customizing open models into their own, and venture firms are encouraging it.
Keep the caveats in view. Tenet is a research preview, and Harvey says the next step is scaling compute to bring the work from research to production, so the 100,000 plus lawyers across 1,300 organizations using Harvey today still ride the existing stack. Bloomberg published neither a post Tenet margin figure nor a breakdown of what still routes to frontier models for the hardest matters.
The plumbing matters too. Harvey built Tenet with Fireworks AI as its research partner, and inference runs on Fireworks infrastructure across the US, EU, and Australia, on weights sitting on disk. Harvey also says it used zero customer data in post training. Meanwhile Thomson Reuters is shipping CoCounsel on Qwen, another Chinese open weight base. The pattern is unmistakable. Serious AI companies are now comfortable building on open weights from anywhere, as long as the economics and the performance check out.
For you, this is the economic plot twist of the AI boom. The app layer learned it can own its margins by owning its models. OpenAI and Anthropic, both preparing for IPOs, now face their best customers becoming their most capable competitors. The open weight era is arriving on a spreadsheet, and the benchmark charts only confirmed what the margins already proved.
Quick answers
What is this story about?
Here is a number that reshaped a company in six months. Harvey, the legal AI startup valued at about 15.6 billion dollars, watched its gross margins swing from about 50 percent at the start of the year to minus 50 percent by June. Token usage jumped twentyfold after a March update to its AI agent tools, and usage based pricing from OpenAI and Anthropic turned surging demand into a cost engine.
Why does this story matter?
For you, this is the economic plot twist of the AI boom. The app layer learned it can own its margins by owning its models. OpenAI and Anthropic, both preparing for IPOs, now face their best customers becoming their most capable competitors. The open weight era is arriving on a spreadsheet, and the benchmark charts only confirmed what the margins already proved.
Sources
- AI Weekly: Harvey moves flagship off frontier labs to Moonshot's Kimi K3
- BigGo Finance: AI startups rush to build their own models
- beri.net: CoCounsel's new model and Harvey's Tenet lineage
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.