The popular legal AI tool, Harvey, has announced that it will be incorporating leading foundation models from Anthropic and Google into its platform, in addition to its existing use of OpenAI’s models, as stated in a recent blog post.
This development is noteworthy because Harvey is one of the most successful early portfolio companies backed by the OpenAI Startup Fund, an investment fund associated with OpenAI that focuses on supporting companies developing products based on AI technologies, primarily those of OpenAI. Although Harvey emphasizes that it is not abandoning OpenAI’s models, but rather expanding its options, this move is still a significant victory for OpenAI’s major competitors.
Harvey was among the first four startups to receive backing from the OpenAI Startup Fund, as announced in December 2022, during the tenure of OpenAI CEO Sam Altman, who was then leading the fund. The other startups in this initial cohort included Descript, Mem, and Speak.
Since then, Harvey has experienced rapid growth and is now valued at $3 billion, as reported in February when it announced a $300 million Series D funding round led by Sequoia, with participation from prominent investors such as Coatue, Kleiner Perkins, and the OpenAI Fund.
Notably, Google’s venture arm, GV, led Harvey’s $100 million Series C funding round in July 2024, with the OpenAI Fund also participating. However, despite this investment, Harvey did not immediately adopt Google’s AI models. GV also participated in Harvey’s Series D funding round.
The decision to expand beyond OpenAI’s models was informed by Harvey’s internally developed benchmark, known as BigLaw, which demonstrated that a variety of foundation models are increasingly capable of performing legal tasks, with some exceling in specific areas. Instead of focusing on training its own models, Harvey opted to leverage high-performing, reasoning foundation models from other vendors, such as Google and Anthropic, available via Amazon’s cloud, and fine-tune them for the legal market.
By utilizing multiple models, Harvey aims to enhance its capabilities, particularly in the development of AI agents. The company notes that using a range of models will facilitate the creation of more sophisticated AI agents tailored to the legal profession.
Harvey’s benchmark, BigLaw, has shown that seven models, including three non-OpenAI models, now outperform the original Harvey system. The company highlights that different foundation models excel in different legal tasks, with Google’s Gemini 2.5 Pro, for instance, performing well in legal drafting but struggling with pre-trial tasks like writing oral arguments due to its limited understanding of complex evidentiary rules.
In contrast, OpenAI’s o3 model performs well in pre-trial tasks, closely followed by Anthropic’s Claude 3.7 Sonnet. These findings underscore the importance of selecting the most appropriate model for specific legal tasks.
In its blog post, Harvey announces that it will join the growing ranks of companies sharing public leaderboards of model benchmark performance, ranking major reasoning models based on their performance in legal tasks. Furthermore, the company will publish research featuring “top lawyers providing nuanced insights into model performance not captured by single-score benchmarks.”
By adopting competitors’ models and sharing benchmarking results, Harvey is increasing the pressure on its backers, including Google, to continuously demonstrate their capabilities. However, OpenAI should not be overly concerned, as it remains a leader in the field of AI, despite the growing complexity and politicization of AI benchmarking.
Harvey CEO Winston Weinberg expressed his appreciation for OpenAI’s support and investment, stating, “We are incredibly fortunate to have OpenAI as an investor in Harvey and key collaborator in our product. And, we are energized to add to our options for customers as we continue
Source Link
