Harvey's Tenet nearly doubles held-out legal wins over its Kimi K3 base
Legal AI's boldest open-weight bet just got its first real receipts — and Washington is already citing them as policy evidence. Plus: US model prices keep sliding.
Harvey published its first benchmark results for Tenet, and the post-trained Kimi K3 model completes almost twice as many held-out legal tasks as the open-weight base it started from. In a research update posted Thursday, the $11 billion legal-AI company reports a 9-percentage-point jump in all-pass rate on held-out tasks from Vals' LAB legal benchmark, plus a 20% improvement on LAB Contracts — good enough for state-of-the-art on contracts and second place overall. The model was trained with Fireworks through asynchronous reinforcement learning on realistic legal assignments, graded against expert rubrics, with reward shaping that favors trajectories burning fewer tokens so cost stayed flat while quality climbed. Harvey says the gains carried over to agent benchmarks the model had never seen during training, and that no customer data was used.
The cost case is the other half of the pitch: Harvey says Tenet hits these scores at less than a quarter of the cost of leading proprietary models, which is what makes running agents continuously across every matter affordable. Last week's launch asked readers to take the claims on faith — we covered the unveiling when the scores were still promises — Harvey launches Tenet, its first in-house legal model. Now the evals exist, and the strategic picture matters more than any single number: a marquee US application company, itself backed by OpenAI, is building its core intelligence on China's flagship open-weight model instead of renting frontier APIs from the labs it used to sit atop. South China Morning Post framed it as a pivot away from US models; export controls did not stop it, because open weights made it trivial.
David Sacks entered the open-model fight waving Harvey's results as Exhibit A, arguing that restricting open weights would only hand China the advantage. The former White House AI and crypto czar, now co-chair of the President's Council of Advisors on Science and Technology, said "restrictions that kneecap open models would do nothing to stop Chinese labs from shipping the next Kimi," according to Benzinga's account of his remarks. Such limits, he argued, would cripple startups like Harvey that build affordable specialized tools on open systems, while closed labs quietly welcome the blow to their competition. The irony is hard to miss: Washington's most prominent open-model defender is citing a product built on a Chinese base — and the intervention lands days after OpenAI flipped to support California's AI safety law, keeping the open-versus-closed fight at full boil.
Several US AI companies have cut usage prices for their large models, a Sina News report said, as the token price war grinds into another week. The report frames the reductions as competition-driven, and they land in a strange squeeze: compute itself keeps getting pricier — Nvidia tells top customers to expect 15%+ price jumps from early 2027. Tokens get cheaper, GPUs get dearer, and the margin increasingly lives in whatever you own in between — which is exactly the logic behind Harvey's move above.
What to watch: whether Tenet ships into customer-facing Harvey workflows this quarter — and whether Sacks's open-model argument survives contact with actual legislation.
If the best US legal AI now builds on a Chinese open model, who exactly are export controls protecting? Tell us in the comments.
