Harvey built a legal software business valued at $11 billion using other companies' artificial intelligence models. On August 18, 2026, it announced that it had trained its own.
The name is Tenet. And the choice of foundation is what makes the case interesting: it wasn't built from scratch, nor was it based on the strongest model on the market. It was trained on top of Kimi K3, the open-weight, low-cost model from Chinese company Moonshot AI — with legal data produced by Harvey itself, drawn from simulated disputes and case files prepared by lawyers hired specifically for that purpose.
The result the company claims to have achieved: performance comparable to the strongest general-purpose models, at open-model cost.
Why this is counterintuitive
The dominant logic in recent years has been simple: whoever has access to the most powerful model wins. Companies competed to use the newest version of the most expensive model.
What Harvey did flips that equation. It took a foundation that was weaker in general capability and made it better at what matters to it — using something no frontier lab has: the archive and judgment of a specific domain.
A generalist model knows a bit of everything, including law. A model trained on the material of a company that lives and breathes law knows the job the way that work actually requires. For the specific task, the second one wins — and costs a fraction as much.
What changes when the cost per task collapses
This is the part that usually gets overlooked, and it's the most important one. Harvey says the low cost makes it practical to run agents continuously across every case.
Notice the change in nature. When each query is expensive, AI gets used at important moments, under human decision: someone decides to ask. When the cost drops enough, it starts running all the time, on everything — reviewing every case, checking every document, flagging what no one asked it to check.
It isn't the same tool used more often. It's a different way of working, one that only exists below a certain price point.
And what about companies that won't train any model at all?
Worth being direct here: training a model, even on an open foundation, requires scale, data, and a team that the vast majority of companies don't have — and Harvey is worth $11 billion. If the takeaway from this case is "I'm going to train my own," it will backfire.
What's worth taking from it is the principle, and it applies at any size:
- The differentiator isn't the model, it's the data only you have. Your contracts, your reports, your service history, the proposals that closed and the ones that didn't. Everyone rents the same model; no one else has your archive.
- Organizing that archive is the investment that survives switching models. The model you use today will be replaced within months. The well-organized material from your operation keeps its value — and improves the output of whatever model comes next.
- There's a middle path. Between "use ChatGPT like everyone else" and "train your own model" lies the range where nearly every company should sit: feeding the model your knowledge base at the moment of the query, instead of hoping it guesses how your company works.
- Ask what the task costs, not what the model costs. It was the drop in cost per task that changed what Harvey could do. That's the useful yardstick, not the subscription price.
It's worth noting the contrast with the other side of the same coin: running an open model on your own infrastructure has a real cost that headlines tend to hide. Harvey's case doesn't contradict that — it shows where the investment pays off: not in hosting the model, but in specializing it with what you already know.
The honest summary
Two caveats. Tenet is not yet in Harvey's product, and no launch date has been announced — the performance described is what the company reports, not something clients are already using. And the comparison numbers come from Harvey itself, without independent verification.
What's already clear, regardless of what happens with Tenet: the company that had the most reason to keep renting other people's models — because it built $11 billion doing exactly that — concluded it was worth having its own, built on a cheap, open foundation, with data only it possesses.
Facts compiled from Harvey's August 18, 2026 announcement, Business Insider's coverage, and August 2026 technology news. Tenet's performance is reported by the company itself.
Perguntas frequentes
What is Harvey Tenet?
It is the first in-house model from Harvey, the legal software company valued at US$ 11 billion, announced on 18 August 2026. It was trained on top of Kimi K3, the low-cost open-weights model from China's Moonshot AI, using Harvey's own legal data drawn from mock disputes and case files produced by lawyers.
Why train on an open model instead of using the strongest model on the market?
Because the comparison that matters is not general capability but performance on the specific task. A generalist model knows a bit of everything, law included; a model trained on the archive of a company that lives on legal work knows it the way that work demands. Harvey says it reached performance comparable to the strongest general models at the cost of an open model.
Why does cost per task matter so much?
Because it changes what is possible, not just what you spend. When each query is expensive, AI gets used at important moments, by human decision. When cost falls far enough, it becomes practical to run agents continuously across every case — reviewing everything, all the time. That is a different way of working, and it only exists below a certain price.
Should my company train its own AI model?
Almost certainly not. Training a model, even on an open base, requires scale, data volume and a team that the vast majority of companies do not have — and Harvey is worth US$ 11 billion. The middle path serves most: feed the model your knowledge base at the moment of the question, rather than hoping it guesses how your company works.
What is the lesson for small and medium companies?
That competitive advantage is not in the model but in the data only your company has: contracts, reports, service history, the proposals that closed and the ones that did not. Everyone rents the same model; nobody has your archive. And organising that archive is the investment that survives a change of model, because the model will be replaced within months and your material will still be worth something.
Is Tenet already in use?
No. The model is not yet available in Harvey's product and no launch date has been announced. The performance described is what the company itself reports, without independent verification — which is reason for caution before treating the figures as settled.