Standard authorized AI device Harvey will now be utilizing main basis fashions from Anthropic and Google, transferring past strictly utilizing OpenAI’s, Harvey introduced in a weblog submit on Tuesday.
That is noteworthy as a result of Harvey is likely one of the OpenAI Startup Fund’s most profitable early-backed portfolio corporations. The OpenAI Startup Fund is an OpenAI-associated fund to again corporations growing merchandise on high of AI applied sciences, primarily OpenAI’s personal. Whereas Harvey says it’s not abandoning OpenAI, merely including extra fashions and clouds, that is nonetheless an enormous coup for OpenAI’s large rivals.
Harvey is likely one of the first 4 startups that the OpenAI Startup Fund backed, it stated in December 2022. That was again when OpenAI CEO Sam Altman was nonetheless working the fund. (The others in that first cohort embody Descript, Mem, and Communicate.)
Harvey, which has grown like mad since then, is now a $3-billion-valuation startup, it stated in February, when it introduced a $300 million Sequence D led by Sequoia, with different large names like Coatue, Kleiner Perkins, and the OpenAI Fund piling in.
Curiously, Google’s enterprise arm, GV, led Harvey’s $100 million Sequence C in July 2024 (and the OpenAI Fund participated in that spherical, too). However Harvey didn’t instantly undertake Google’s AI fashions after it put Google’s company enterprise agency on its cap desk. (GV participated in Harvey’s Sequence D as effectively.)
So, what satisfied Harvey to maneuver past OpenAI’s fashions now? The startup’s internally developed benchmark, dubbed BigLaw, confirmed that all kinds of basis fashions are rising more and more adept at a variety of authorized duties and a few are higher at particular duties than others.
As a substitute of spending its efforts coaching fashions, Harvey figured, it may merely embrace high-performing, reasoning basis fashions from different distributors (e.g. Google and Anthropic by way of Amazon’s cloud) after which fine-tune them for the authorized market.
Utilizing a wide range of fashions may even assist as Harvey creates AI brokers, the corporate says.
“In less than a year, seven models (including three non-OAI models) now outperform the originally benchmarked Harvey system on BigLaw Bench,” Harvey wrote within the weblog submit.
Harvey’s benchmark additionally confirmed that totally different basis fashions are higher at particular authorized duties than others. As an example, it says Google’s Gemini 2.5 Professional “excels” at authorized drafting however “struggles” with pre-trial duties like writing oral arguments as a result of the mannequin doesn’t absolutely perceive “complex evidentiary rules like hearsay.”
OpenAI’s o3 does such pre-trial duties effectively, in keeping with Harvey’s testing, with Anthropic’s Claude 3.7 Sonnet following shut behind.
In its weblog submit, Harvey says that it’ll additionally now be part of the rising ranks of these sharing a public leaderboard of mannequin benchmark efficiency. Its board will rank how main reasoning fashions are doing on authorized duties. And the corporate gained’t simply boil rankings all the way down to a single quantity, however may even publish analysis the place “top lawyers provide nuanced insights into model performance not captured by single-score benchmarks.”
So, not solely is OpenAI-backed Harvey adopting rivals’ fashions, it’s also amping up the strain on its backers (together with Google) to maintain proving themselves. Not that OpenAI ought to fear a lot on that rating. Whereas AI benchmarking is rising more and more complicated and considerably political, it is a world the place OpenAI nonetheless shines.
“We are incredibly fortunate to have OpenAI as an investor in Harvey and key collaborator in our product,” Harvey CEO Winston Weinberg informed TechCrunch in a press release. “And, we are energized to add to our options for customers as we continue to serve the needs of our customers globally.”
