OpenAI Is Moving Into Legal AI. Harvey Has to Keep Moving Too.

Astra for Law raises the baseline. Harvey's same-day research points to the harder contest: whose system learns to do work lawyers will keep paying for.

Richard TangRichard Tangfour-minute read
Share
The glass facade of 1515 Third Street, OpenAI's headquarters building in San Francisco's Mission Bay neighborhood, seen from the corner of Warriors Way.
OpenAI’s headquarters at 1515 Third Street in San Francisco, photographed June 15, 2025. Photo: Coolcaesar / Wikimedia Commons (https://commons.wikimedia.org/wiki/File:1515_Third_Street.jpg), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

OpenAI$1.18T — OpenAI, private, latest valuation $1.18T introduced Astra for Law on September 17, combining GPT-6 Astra with legal search and specialized instructions. Harvey$15.5B — Harvey, private, latest valuation $15.5B and Legora$5.55B — Legora, private, latest valuation $5.55B are named as future API users; selected firms get initial access through ChatGPT and Codex. The supplier is also offering a direct route to the customer. 1

The threat is serious. It does not mean Harvey's business has just disappeared.

That conclusion assumes Harvey stands still while OpenAI advances. The useful question is what Harvey can learn and build as capabilities that once required specialist engineering become available from its supplier. A better foundation can strengthen a product and make parts of it harder to charge for at the same time.

There is a concrete reason to ask that question today. Alongside OpenAI's launch, Harvey published research on using lawyers' judgments to evaluate and improve legal AI. Its research exposes a difficult constraint: getting an answer judged good enough for a particular task. 2

The baseline has moved. The spending is still unknown.

OpenAI's index spans more than 230 million URLs. The announcement describes it as complementary to licensed content from providers including Thomson Reuters. It supplies no price, paid-seat total, or revenue for Astra for Law. It also does not establish that using the index is free. 1

Those omissions matter. A large corpus tells us about the material a system can search. It does not tell us what a firm can cancel, what it will pay for the replacement, or whether the replacement completes the same work.

Our earlier Harvey piece examined the company's financing and valuation. 3 Harvey's own September 9 announcement reported a $550 million raise at a $15.5 billion valuation, and said 80% of Am Law 100 firms used its products. Those are company-reported figures; usage does not tell us the size or profitability of those relationships. 4

The new launch creates a reason to revisit that business. It does not give us a new revenue multiple or evidence that customers are leaving.

The difficult part is deciding what better means

Harvey describes a study covering 24 legal-agent tasks, with three lawyers rating each task. Reviewers were unanimous in only 25% of matchups. The company says much of the disagreement concerned how to weigh the same strengths and weaknesses. 2

This is where the problem gets interesting. Imagine two draft answers: one is easier to read, while another handles an important exception more carefully. Which should the system learn to produce? Counting positive ratings without understanding the reason could teach it the wrong lesson.

Harvey is experimenting with generative reward models: AI evaluators that compare outputs against criteria shaped by legal experts. It reports that these evaluators sometimes accepted serious errors in exchange for other strengths, or selected a winner when lawyers considered both answers inadequate. These are findings from Harvey's own research, not independent validation. 2

A system can produce more answers without producing a reliable signal about which answers deserve to be repeated. That bottleneck deserves scrutiny before anyone celebrates a data flywheel.

Conceptual legal AI learning loop: do the task, review the result, keep useful feedback, then improve and retest. Keeping useful feedback is highlighted as the bottleneck.
Where the learning loop can break. Conceptual illustration of this article’s argument, informed by Harvey’s research [2]; not a verified architecture of either company. Original diagram by Codex for The Inference; created for this article.

The work around the model matters: what the evaluator can inspect, how a lawyer's correction is captured, what counts as a serious failure, and whether the next version improves on unfamiliar work. More usage would only be an advantage if the company could turn the relevant experience into dependable improvement, within the permissions governing that information.

Having customers is not, by itself, evidence that this loop exists. Showing the loop improve customer outcomes would be much more persuasive than describing the product as specialized.

OpenAI can keep learning too

In OpenAI's own test on 200 private validation questions, Astra for Law passed the overall correctness check on 54.0%, versus 38.7% for Astra using web search. OpenAI also describes firm-specific applications built with Sullivan & Cromwell, Ropes & Gray, and Cooley. 1

The benchmark is a narrow comparison, not a measure of completed client matters. But the firm projects make the competitive question harder. It would be a mistake to assume the foundation provider must remain distant from the work while specialists alone understand customers.

Harvey cannot rely on today's division of labor remaining intact. Nor should we assume OpenAI's launch is the final product against which Harvey will compete. Both sides can improve.

The specialist's best move is to use the stronger foundation where it helps and concentrate effort on the difficult work customers still cannot complete satisfactorily. The evidence to look for is fewer consequential corrections, less review effort, and repeat use on meaningful tasks. Those are proposed tests, not results either announcement establishes.

The next commercial evidence should be equally specific: a renewal, a procurement decision, disclosed pricing, or a customer explaining which work moved and why. Continued payment would tell us customers still found value. We would then need to understand what they were buying, rather than deciding on their behalf that a general-purpose alternative ought to be enough.

Astra for Law makes standing still more dangerous. It does not make building a specialist pointless. Harvey's opportunity is to turn a stronger foundation into a better working product faster than the foundation provider can close that remaining gap. The company will have to keep earning that gap. Its valuation cannot do the work for it.

Get Applications by email, free

How AI is changing the software people and companies use. In our daily email, every weekday.

Reader comments

0 comments

    More in Applications

    All in this topic →

    Applications

    Four days after reaching ChatGPT's 1.2 billion users, Synthesia capped its free video tier at one a month

    OpenAI killed Sora in March, and a Synthesia plugin now fills the video slot inside ChatGPT, in front of the chatbot's 1.2 billion weekly users. Four days after launch, the $4 billion London company capped its free tier at one AI video a month.

    Applications

    Omnicom's creative networks lose leaders as $1 billion of its Interpublic merger savings come out of labor

    Omnicom's third quarter showed 7.2% organic growth and merger savings arriving on schedule, yet about $1 billion of the $1.5 billion program comes out of labor, and the labor works in the creative shops now shedding executives. December's quarter is the first test of both claims at once, and the answer reads through to Publicis and WPP.

    Applications

    Adobe will let rival AI agents into Photoshop only on terms that keep Adobe paid

    Adobe will host outside agents inside its apps while rival usage is billed on someone else's account, and its own research says 58% of the companies it talks to are stuck in an AI pilot or unable to deploy at scale. The December 9 report, the first under new CEO Anil Chakravarthy, tests which of the two explanations for Adobe's falling net-new ARR holds.

    Applications

    Chegg lost 40 percent of its revenue and stopped reporting subscribers

    AI wiped out the homework-help business and now builds its replacement cheaply. But the pivot is measured in site visits, the subscriber count has vanished, and a stock at $0.61 now faces two listing clocks.

    Applications

    October's record container imports flow straight into Descartes' December 2 quarter

    Descartes Systems Group bills per transaction on multiyear contracts, so the record October container volume its own report measured lands directly in the quarter it reports on December 2. The same surge is the risk if tariff front-loading unwinds.

    Applications

    Coursera trades just $113 million above its cash pile while its own buyback sits $71 million underwater

    The market pays about $113 million, under one times forward EBITDA, for a business guiding to $1.23 billion of revenue, because the merger math fails and management's own $215 million buyback sits $71 million underwater. A 14.7% holder and a poison pill set the collision date.

    Subscribe to TheInference for free