Harvey has quietly stopped renting the AI it runs its legal business on.
The company launched Tenet, its first in-house model purpose-built for legal work. Tenet is post-trained on Kimi K2, an open-weight model from the Chinese startup Moonshot, rather than built from scratch. It replaces OpenAI's models inside Harvey's own product, even though OpenAI is one of Harvey's backers. That makes Tenet less a research showcase and more a supply-chain move.
Every AI-native startup eventually confronts the same math: margins shrink when your core product runs on someone else's metered API. Post-training an open-weight base instead of training a foundation model from scratch is the cheaper, faster path to that independence. It also means a company partly funded by OpenAI is now building tech to replace OpenAI's models in its own flagship product - an awkward but increasingly common tension in the AI industry.
Harvey hasn't published benchmarks showing Tenet beats the OpenAI models it's replacing, so for now this looks like a cost and control story first, a quality story second.