Harvey Launches Its Own AI Model — and a Direct Challenge to OpenAI and Anthropic

Harvey, the legal AI startup valued at $11 billion, has unveiled its first proprietary model called Tenet — a significant strategic pivot for a company that built its business entirely on top of models from OpenAI and Anthropic. The move is a direct response to an increasingly uncomfortable reality: the very AI companies Harvey depends on are now chasing its customers. Anthropic has released legal-focused plugins for document review and drafting, while OpenAI hired Ironclad founder Jason Boehmig to lead a dedicated legal push. Google and Meta are believed to be eyeing the space too.

Tenet is built on Kimi K3, a low-cost, open-source model from Chinese startup Moonshot AI that caused a stir when it launched in July 2025 for its combination of power and affordability. Harvey's team — led by co-founder Gabe Pereyra, a former Google DeepMind researcher — trained Kimi K3 using synthetic legal data generated by staff and contract attorneys hired through platforms like Mercor and Snorkel. These lawyers invented mock disputes and case files, then scored how well models reasoned through them, producing training material calibrated to actual legal thinking rather than general-purpose benchmarks.

The economics are central to the play. Every query a lawyer runs through Harvey's platform costs the company a fee to its upstream model provider. Routing more of that traffic through a proprietary engine could meaningfully improve margins without raising prices. But Harvey frames Tenet as a quality play as well — another specialized option in the mix of models Harvey already routes tasks through based on their respective strengths.

The more ambitious goal, though, is downstream: Pereyra envisions Tenet eventually becoming a foundation model that individual law firms fine-tune on their own historical work — decades of negotiated contracts, litigation strategies, and deal terms currently locked inside lawyers' heads or buried in archives. Under this model, Harvey would supply the base and help clients build bespoke versions on top, shifting the company's identity from software vendor toward something resembling a professional services firm that helps clients operationalize their institutional knowledge.

Tenet is not yet live in Harvey's product, and the company hasn't disclosed a launch timeline or named any law firms currently testing it. Harvey has promised to publish benchmark results comparing Tenet against rival models, though those figures warrant scrutiny — model developers routinely train on the same benchmarks they use to measure themselves, which inflates scores over time. The broader rollout, dubbed Harvey II, also includes a new Memory feature that lets users store working preferences so Harvey's agents can apply them consistently across tasks.