Harvey Reached $11B Before Its In-House Model

By August 2026, Harvey was valued at $11 billion after building its legal-software business on models it did not own. Three years earlier, investors valued the company at $715 million and reporters defined it by its use of OpenAI technology. That increase came before Harvey launched an in-house legal model.

No disclosed pricing analysis explains how investors apportioned the $11 billion among Harvey’s software, customers, growth, and model strategy. The observable sequence is narrower: Harvey accumulated revenue, customers, and enterprise value before bringing part of the model layer in-house.

Harvey built distribution before it built a model

In December 2023, Harvey raised an $80 million Series B while building legal tools with OpenAI technology. Law firms and legal teams bought that product before Harvey owned foundation-model intellectual property.

By February 2025, Harvey said annual recurring revenue had surpassed $50 million after it raised $300 million at a $3 billion valuation. In June 2025, the company said it served 337 clients, including KKR and PwC, across 53 countries. By August 2026, Harvey’s legal-software business was valued at $11 billion.

Harvey had secured distribution, customer relationships, and domain usage before launching Tenet. The model expanded a commercial position that already existed.

That sequence supports a limited vertical-AI claim: a vendor can build enterprise distribution on rented foundation models and later add direct control over adaptation and evaluation. Tenet may also affect latency, cost, and deployment choices, but Harvey has not reported realized improvements on those measures.

Tenet makes the matter larger than the prompt

Harvey Tenet is the company’s first proprietary in-house model for legal work. Harvey trained it on mock disputes and case files using an adaptation of Kimi K3. The training material gives Tenet legal specialization; Harvey II supplies the surrounding system.

Harvey II agents begin with the files, context, permissions, and history attached to a matter or project. Users can assign tasks to lawyers or agents and track the work. The design unit expands from a single request to a continuing body of work.

Tenet shapes the answer, while Harvey II controls what the agent may see, who acts next, and how the system records completion.

A lawyer can work from the matter instead of reconstructing it in every conversation. Harvey II carries forward the approved files, permissions, history, and task assignments. Those controls reduce repeated setup; they do not guarantee correct authorities or adequate review.

Tenet buys control while preserving upstream dependence

Moonshot announced Kimi K3 on July 17, 2026, and said it planned to release the full model weights by July 27. Harvey unveiled Tenet the following month as an adaptation of Kimi K3, gaining a legal model without reproducing Moonshot’s original pretraining run.

That route leaves Harvey dependent on an upstream architecture even as it brings legal specialization in-house. Future Kimi releases, alternative open-weight models, and hosted systems can still change the cost and performance of Harvey’s model choices.

Operating an open-weight model also carries serving costs. Documents reported in April 2026 showed that OpenAI and Anthropic projected inference costs above half of revenue. Harvey has not disclosed comparable unit economics, so the frontier-lab figures cannot be transferred directly to Tenet. They make serving cost a live procurement constraint.

Harvey must compare hosted and self-operated models as performance, privacy, deployment control, and unit costs move. Tenet gives the company more control over legal evaluation and deployment, while Harvey II can route different tasks through different systems. The relevant measure is cost per accepted task, including retrieval, review, and correction.

Legal value depends on accountable completion

A 2023 New York sanctions case exposed the gap between fluent output and completed legal work. A lawyer faced sanctions after a ChatGPT-assisted filing included bogus judicial decisions, quotes, and citations.

The sanctions order held the lawyers before the court responsible for the filing. A usable legal system therefore has to retrieve and apply the right material, preserve access boundaries, show its sources, and route uncertain work to someone authorized to judge it.

Garfield shows how AI can enter a regulated legal service under narrower conditions. The UK-based AI law firm received regulatory approval in 2025 and won its first case in the English courts in June 2026. The unpaid-fees matter involved a freelancer who paid about £400 for technology to draft documents for a £7,000 claim. One small case does not establish broad reliability, but it does show AI-assisted work reaching judgment through an accountable firm.

In both the sanctions case and Garfield’s claim, an identified lawyer or regulated firm remained in the chain. Human review consumes time, but it also determines who verifies the authorities, resolves ambiguity, and stands behind the result for the client and the court.

Benchmark scores omit system risk

Anthropic reported that Opus 4.6 scored 90.2% on BigLaw Bench, its highest result for a Claude model. The benchmark measures legal capability under defined tasks. A firm must still decide which client files the model may access, whether its sources are authoritative, how to route the draft, and who reviews the final work.

Harvey needs model evaluations to decide whether Tenet, an OpenAI model, an Anthropic model, or another option should handle a task. A general benchmark can answer how well a model performed on the test. It cannot determine whether a system completed permitted work to a firm’s professional standard.

Each layer can fail independently. The model can reason badly, retrieval can supply the wrong authority, permissions can expose the wrong matter, orchestration can send work to the wrong person, and a lawyer can approve a defective result. One model score leaves those system risks unmeasured.

Information rights keep incumbents in the contest

AI-native vendors such as Harvey and Legora begin with work surfaces and move deeper into models. Thomson Reuters and LexisNexis begin with legal information products and established workflows. Each side controls a different part of the system.

Information rights impose a concrete limit on model development. In 2025, Thomson Reuters defeated Ross Intelligence’s fair-use defense in a major US copyright ruling. The decision shows that using protected legal content for model development can bring licensing and litigation costs.

Harvey still needs reliable rights, sources, and evaluations. Thomson Reuters and LexisNexis still need products that turn their information into completed work. Anthropic and OpenAI can supply reasoning engines, while legal customers supply matter permissions, institutional rules, and review chains.

Harvey and its rivals are competing across five control points: model selection, domain evaluation, matter context, information rights, and accountable review. Owning one leaves a vendor dependent on the owners of the others. A vendor that assembles all five can change models with less disruption to how its customers complete and approve legal work.

Harvey’s $11 billion now reads differently. The valuation preceded Tenet, although no public evidence shows that investors specifically priced matter-level control. Harvey II makes the product strategy visible by placing generation inside files, permissions, routing, and review—the place where the next sentence becomes someone’s responsibility.