Harvey embraces open AI models to reduce reliance on Anthropic and OpenAI

Harvey built its legal AI platform on the backs of OpenAI and Anthropic. Then the bill arrived. The company, which provides AI tools to law firms and corporate legal departments, watched its gross margins collapse from roughly 50% to negative 50% within six months of a March 2026 product update that triggered a sharp surge in customer usage. More customers, more queries, more inference costs from proprietary providers. The math stopped working fast. Harvey’s response: build its own model and reduce the power that external AI providers hold over its cost structure. The company launched Tenet, its first internally post-trained open-weight model, around August 20, 2026. It is built on Kimi K3, an open-weight base model with approximately 2.8 trillion parameters that was itself open-sourced around July 2026. From cost crisis to custom model Harvey apparently moved quickly. Going multi-model is not new for the company: it integrated Anthropic and Google models alongside OpenAI back in May 2025, creating a layered architecture that could route tasks to different providers. But routing among expensive proprietary models only helps so much when all of them charge meaningfully for inference...