The popular legal AI tool Harvey will use the main foundational models of Anthropic and Google and will move using Openai strictly, Harvey announced in a blog post Tuesday.
This is noteworthy as Harvey is one of the most successful early portfolio companies in the Openai Startup Fund. The Openai Startup Fund is an Openai-related fund primarily for back companies that develop products in addition to Openai’s proprietary AI technology. Harvey has not abandoned Openai, it’s not just adding models and clouds, but this is still a big coup for Openai’s large competitors.
Harvey said in December 2022 that it was one of the first four startups supported by the Openai Startup Fund. This was when Openai CEO Sam Altman was still running the fund. (Others in that first cohort include writing, MEM, and speaking.)
Having grown crazy ever since, Harvey is now a $3 billion startup, he said in February when he unveiled a $300 million Series D, which has been stacked with other big names like Cotue, Kleiner Perkins and Openai Fund.
Interestingly, Google’s venture arm GV led Harvey’s $100 million Series C in July 2024 (and the Openai fund also joined the round). However, Harvey did not immediately adopt Google’s AI model after placing Google’s corporate venture company on the cap table. (GV also participated in Harvey’s Series D.)
So, what persuaded Harvey to move beyond the currently open model? A benchmark called Biglaw, a benchmark developed internally by startups, has shown that various underlying models are increasingly proficient in a variety of legal tasks, with some being better at certain tasks than others.
Instead of spending effort training models, Harvey thought they would simply embrace high-performance, inference foundation models from other vendors (such as Google and humanity via Amazon’s cloud) and tweak them in the legal market.
As Harvey creates AI agents, it also helps to use a variety of models, the company says.
“In less than a year, seven models (including three non-OAI models) have surpassed the Harvey System, originally benchmarked on the Biglaw bench,” Harvey wrote in a blog post.
Harvey’s benchmarks showed that different underlying models were superior to others in certain legal tasks. For example, Google’s Gemini 2.5 Pro says it is “good” at legal drafting, but “struggles” with pre-trial tasks such as writing oral arguments. This is because the model does not fully understand “complex rules of evidence like hearsay.”
According to Harvey’s tests, Openai’s O3 does such pre-trial tasks well, followed by Anthropic’s Claude 3.7 Sonnet.
In a blog post, Harvey also said he will be taking part in the growth ranks of people who share the public leaderboards of model benchmark performance. Its board ranks how the key inference models are doing in legal tasks. And the company will not only boil the rankings into a single number, but also publish a survey that “provides subtle insights into model performance that top lawyers have not been captured in single score benchmarks.”
So, Openai-backed Harvey not only adopts a competitor model, it also increases the pressure on supporters (including Google) to continue to prove themselves. Openai doesn’t have to worry too much about its score. The AI benchmark has become increasingly complicated and somewhat political, but this is a world where Openai still shines.
“We are extremely fortunate to have Openai as a Harvey investor and a major collaborator of our products,” Harvey CEO Winston Weinberg told TechCrunch in a statement. “And now we’re adding options for our customers as we continue to meet our customers’ needs globally.”