A highly anticipated legal artificial intelligence tool, Harvey, recently announced a strategic adjustment, planning to integrate leading foundation models from Anthropic and Google, marking the end of its previous reliance solely on OpenAI technology. Notably, Harvey was one of the star investment projects in OpenAI's startup fund, which aimed to support companies based on AI technology, particularly OpenAI’s proprietary technology. Although Harvey emphasized that it would not abandon OpenAI but merely add more model and cloud service options, this move is undoubtedly a significant victory for OpenAI's competitors.
In December 2022, OpenAI’s startup fund listed Harvey as one of its first four supported startups, at a time when the fund was managed by OpenAI CEO Sam Altman. Since then, Harvey has developed rapidly, with its valuation reaching $3 billion. In February this year, the company announced a $300 million Series D financing round led by Sequoia Capital, with multiple well-known institutions, including Coatue, Kleiner Perkins, and the OpenAI Fund, participating.
Image Source Note: The image is generated by AI, licensed by MidJourney
Notably, in July 2024, Google’s venture capital arm, GV, led Harvey's $100 million Series C funding round, with the OpenAI Fund also participating. However, despite Google’s corporate venture capital department becoming a shareholder in Harvey, Harvey did not immediately adopt Google’s AI models.
The key factor driving Harvey's strategy shift lies in its internally developed benchmark, BigLaw. This benchmark shows that various foundation models are increasingly proficient in legal tasks, and certain models outperform others in specific tasks. Harvey believes that instead of investing substantial resources in training models, it can directly adopt high-performance inference foundation models from other suppliers, such as those provided through Amazon Cloud, and fine-tune them for the legal market. The company stated that using diverse models will help build more powerful AI agents.
In a blog post, Harvey noted: "Within less than a year, seven models (including three non-OpenAI models) have outperformed the initial benchmark Harvey system on the BigLaw Bench."
Harvey’s benchmark test results revealed the advantages of different models in specific legal tasks. For example, Google’s Gemini2.5Pro performs exceptionally well in drafting legal documents but falls short in pre-trial tasks like writing oral arguments due to its limited understanding of complex evidence rules such as hearsay. According to Harvey’s tests, OpenAI’s o3 performs well in these pre-trial tasks, while Anthropic’s Claude3.7Sonnet follows closely behind.
Harvey announced that it will join the growing list of publicly available model performance rankings, ranking mainstream inference models’ performance in legal tasks. More importantly, the company will not only release a single ranking score but also share "detailed insights from top lawyers about model performance, which cannot be captured by a single score benchmark."
Therefore, Harvey, supported by OpenAI, not only began adopting rival models but also imposed higher performance requirements on its supporters, including Google. Despite the increasing complexity and competitiveness of AI benchmarks, OpenAI continues to maintain strong competitiveness in this field.
In a statement, Harvey CEO Winston Weinberg said: "We are very honored that OpenAI is an investor in Harvey and a key partner in our products. We will continue to meet the needs of global customers and strive to provide more choices for our clients."