Popular Legal AI Tool Harvey will now be using anthropic and leading Foundation model from Google, moving strictly using Openi, Harvey Announced In a blog post on Tuesday.
This is notable because Harvey is one of the most successful initial-supported portfolio companies of Openai Startup Fund. Openai Startup Fund AI is an Openai-connected fund to refund companies that develop products on top of technologies, mainly Openai’s own. While Harvey says that it is not leaving Openai, only adding more models and clouds, it is still a very large coup for Openai’s big contestants.
Harvey is one of the first four startups that supports the OpenAI Startup Fund, it said In December 2022This was when OpenEE CEO Sam Altman was still running funds. (Others in that first coort include descript, memes and speech.)
Harvey, which has grown like crazy since, is now $ 3 billion-valence startup, it was said in February, when it Announced With other big names such as Cutu, Clener Perkins, and OpenEE Fund Piling, a $ 300 million series D.
Interestingly, Google’s Venture Arm, GV, led Harvey’s $ 100 million series C in July 2024 (and Openai Fund also participated in that era). But Harvey did not immediately adopt Google’s AI model after Putting Google’s corporate venture firm on his cap table. (GV also participated in a series D of Harvey.)
So, what did Harvey assure to move beyond Openi’s model? The internally developed benchmarks of the startup, dubbed deterioration, showed that a variety of foundations models are growing rapidly in a range of legal functions and some are better in specific functions than others.
Instead of spending the training model of their efforts, Harvey felt, it can only embrace the argument, and then fix them to the legal market from other vendors (such as anthropic through the cloud of other vendors (such as Google and Amazon).
Using different types of models will also help because Harvey makes AI agents, the company says.
Harvey wrote in a blog post, “In less than a year, seven models (including three non-OAI models) now perform better than the horny system originally benchmarking on the deteriorating bench.”
Harvey’s benchmark also showed that separate Foundation models are better in specific legal functions than others. For example, it states that Google’s Gemini 2.5 Pro “Excel” in legal drafting but “struggle” such as “struggle” to write oral arguments with pre-examination tasks because the model does not fully understand the complex clear rules such as “Haras”.
According to Harvey’s trial, the O 3 of Openi performs such pre-trial functions well, behind the cloud 3.7 sonnet of anthropic.
In his blog post, Harvey says that it will now also be included in the growing rank of those who share a public leaderboard of model benchmark performance. Its board will rank how the major arguments on legal tasks are models. And the company will not only boil the ranking to the same number, but will also publish research where “top lawyers provide fine insight into the model performance captured by single-score benchmark.”
Therefore, not only Openai-supported Harvey to adopt the model of contestants, it is also increasing the pressure on its backers (including Google) to prove themselves. It is not that Openai should worry more on that score. While the AI benchmarking is becoming increasingly complicated and somewhat political, it is a world where openi still shines.
“We are incredibly lucky to have an investor in Harvey as an investor in our product and a major ally in our product,” Harvey CEO Winston Wenberg told Techcrunch in a statement. “And, we are active for customers to add our options because we continue to meet the needs of our customers globally.”
