Harvey, a prominent legal-focused AI platform previously backed exclusively by OpenAI, announced this week that it would integrate advanced foundation models from Anthropic and Google into its system. This move signals a notable expansion beyond its original reliance on OpenAI’s technology, although the company insists it is not abandoning its initial partner; instead, it is broadening the variety of models available to its users.
The decision is a significant milestone, given Harvey’s high profile within OpenAI’s Startup Fund. Launched as one of the first companies endorsed through OpenAI’s investment arm, Harvey quickly rose to prominence as an exemplary success story among the fund’s earliest bets. Having secured backing from the OpenAI Startup Fund in December of 2022, the startup drew attention due to its rapid growth trajectory, eventually reaching a valuation of approximately $3 billion during a highly publicized Series D funding round in February. Sequoia Capital led that round, joined by notable investors including Coatue, Kleiner Perkins, and the OpenAI Fund itself.
Interestingly, Google’s own venture arm, GV, spearheaded Harvey’s earlier $100 million Series C round in July 2024, although Harvey did not immediately transition to Google’s AI models at the time. GV also participated in Harvey’s subsequent Series D funding.
Harvey’s decision to diversify its foundation model providers stems from internal benchmarking results through its proprietary BigLaw benchmark. This evaluation clearly showed certain models excelling in specific legal contexts, convincing Harvey executives that adopting multiple state-of-the-art models would allow them to fine-tune tools more effectively for targeted legal applications. Harvey noted that Google’s Gemini 2.5 Pro exceeded expectations in legal drafting while falling short on complex pre-trial tasks, such as constructing oral arguments due to limitations in handling nuanced evidentiary issues like hearsay. By contrast, models from OpenAI—particularly the o3 variant—performed notably better on such pre-trial procedures, followed closely by Anthropic’s Claude 3.7 Sonnet.
Additionally, Harvey indicated that relying on a broader set of models would significantly boost capabilities as it develops more sophisticated AI-driven agents for the legal industry. The company observed that in fewer than twelve months, seven different models, including three not based on OpenAI technology, surpassed Harvey’s original system on its benchmarks for legal reasoning tasks.
As part of this shift, Harvey has also committed to publicly sharing the performance metrics from its internal benchmarking process. It will create a publicly accessible leaderboard measuring AI reasoning models against various legal tasks, providing not only numerical rankings but also expert commentary and insights contributed by leading legal professionals. This nuanced and transparent approach is intended to hold AI providers accountable and foster continued refinement of performance across the industry.
Despite this broader embrace of competing models, Harvey’s leadership expressed continued appreciation for their original partnership. “We are incredibly fortunate to have OpenAI as an investor in Harvey and as a key collaborator,” said Harvey CEO Winston Weinberg in a statement. He underscored the excitement about adding new models into their offerings as Harvey remains committed to meeting global customer needs.
