Harvey, the legal AI startup valued at $15.5 billion, has built and launched Tenet, its first proprietary AI model, to reduce reliance on expensive third-party models like OpenAI and Anthropic. The move reveals a fundamental shift in how well-funded AI companies are thinking about model costs, data ownership, and competitive advantage. Rather than paying per token to rent intelligence from frontier model labs, Harvey trained Tenet on attorney-generated case files and legal expertise, then embedded it into Harvey II, a broader platform relaunch announced on August 18.

What Makes Harvey's Tenet Different From General AI Models?

Tenet is not a general-purpose chatbot wearing a lawyer's jacket. It is purpose-built for the grinding, long-horizon work that defines legal practice. Harvey trained Tenet on a customized base from Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model, then post-trained it using Fireworks AI for tasks like contract reviews, due diligence runs, evidence sweeps, and litigation work that can consume hours and hundreds of thousands of tokens per task.

The training data is where the real competitive advantage lives. Harvey combined synthetic data, publicly available legal information, and human expert datasets designed to simulate how lawyers actually work. The company worked with Mercor and other partners to build and validate these expert datasets. Harvey's training environments gave Tenet a short request from a partner, a client matter with relevant files, and an expert rubric describing what good legal work had to contain. Some training runs exceeded 1,000 turns and consumed hundreds of thousands of tokens, mimicking the slog of real legal practice.

The results are measurable. Harvey says post-training helped Tenet complete almost twice as many held-out LAB tasks as the base Kimi K3 model and 20% more LAB Contracts tasks, with all-pass rates rising by 9 percentage points and 2 percentage points respectively. Tenet also reached first place on LAB Contracts and second place on LAB overall, Harvey's internal benchmarks.

How Does Cost Drive the Business Logic Behind Tenet?

Cost is the sharpest point of this story. Harvey says Tenet runs at less than a quarter of the cost of leading foundation models while maintaining strong legal benchmark performance. For a legal AI platform, this matters enormously. Legal work is not a single prompt and response. It is a marathon of long contracts, prior filings, diligence folders, emails, and internal notes. Every token sent to a frontier model carries a bill, and those bills add up fast across a law firm's entire docket.

If Harvey can route more legal work through a model it controls, the margin on each matter its software touches improves dramatically. It also gives Harvey a cleaner product to sell than access to someone else's model with a legal workflow wrapped around it. The company is not just offering a tool; it is offering a cost structure that makes AI affordable at scale for large law firms.

Steps to Understanding Harvey's Broader Platform Strategy

• Firm-Specific Training: Harvey is pointing Tenet toward firm-specific customization, allowing law firms to build specialized models and own more of their own intelligence. Two firms using Harvey should end up with different models and different outputs because their work has shaped those systems differently.
• Memory Across Tools: Harvey II adds a Memory feature that learns how individual lawyers structure summaries, cite material, write, and edit, then carries those preferences across Harvey, Microsoft Word, and Outlook. Users can see, change, or turn off what Harvey remembers, and the company says that memory is never used to train its models.
• Secure Matter Spaces: The platform introduces Spaces, which let agents open inside a matter or project with documents, parties, permissions, tasks, and work history already attached. Client data stays inside the relevant Space, and permissions and ethical walls are built into the plumbing.

These features address a core pain point in legal software. Legal work lives or dies on dry plumbing. If permissions and ethical walls are wrong, the cleverest model is useless. Harvey is betting that firms will value a system that learns their preferences, protects their data, and costs a quarter of what frontier models charge.

Why Does Harvey's Choice of Kimi K3 Matter So Much?

The choice of Kimi K3 as Tenet's base is the detail worth sitting with. A well-funded American legal AI company, backed by Sequoia and GIC, is building core legal infrastructure on an open-weight model from China's Moonshot AI rather than paying OpenAI or Anthropic for every token. This is not a story about ideology. It is a story about what happens once open-weight models get good enough that a company with real legal data and real usage would rather own more of its stack.

Moonshot AI closed a funding round at a $31.5 billion valuation in late July and is already in talks to raise again at up to $50 billion ahead of a Hong Kong IPO. Kimi K3 shipped inside Cursor on launch day and its pricing undercuts OpenAI and Anthropic by up to 9 times, accelerating adoption among developers and enterprises looking to reduce model costs.

What Does This Signal About the Broader Legal AI Market?

Harvey is not alone in this shift. TechCrunch reported in April that rival Legora reached a $5.6 billion valuation after adding $50 million to its Series D round, and Legora says its customers include more than 1,000 law firms and in-house legal teams across 50 markets. The two companies are now fighting over the same prize: legal AI that is fast enough and accurate enough, and cheap enough to become normal at large firms.

Bloomberg Law also reported this week that legal tech companies are moving away from pure reliance on OpenAI and Anthropic as they try to control model costs. The pattern is clear. As open-weight models improve and inference becomes cheaper, companies with domain expertise and real usage data are choosing to build proprietary models rather than rent from frontier labs.

Tenet has not yet proved that it can carry the full weight of high-end legal work in the wild. Harvey's own numbers are promising, not final. But the business logic is already visible. The legal AI winners will not only be the companies with the best interface. They will be the ones that know which intelligence to rent and which intelligence to build, and when the bill from the model labs has become too big to ignore.