Harvey and the cost of useful AI

Harvey's recent model work deserves a more serious reading than the claim that it has abandoned good AI to protect its margins. The useful question for a customer is what happens to the quality and cost of their work when the supplier changes the model underneath it.

Bloomberg reported on September 21 that Harvey's gross margin fell from roughly 50% early this year to minus 50% in June. It subsequently returned to positive territory after an August model launch and other changes. The financial figures came from unnamed sources; Harvey declined to discuss them.

That is a striking account of the cost of serving demand. It is also a reason to examine how an AI product earns its subscription, rather than assume the most expensive underlying model is always the best purchase.

What Harvey actually built

Harvey's August research describes Tenet, a model post-trained on Moonshot's Kimi K3 with Fireworks. The company reports leading performance on LAB Contracts and second place on LAB. Those are Harvey's reported benchmark results, not independent evidence of performance on a particular firm's matters.

Its method includes rewarding more efficient reasoning and tool use while preserving task performance. That matters because buying cheaper tokens is only one way to reduce a bill. Avoiding unnecessary work is another.

None of this establishes that every Harvey task now uses Tenet, or that the change made the product worse. A model's licence and country of origin do not provide a score for its output. Buyers need task-specific evidence. Calling something a frontier model does not remove that requirement either.

Our view is that the model choice is a legitimate engineering decision. The commercial test is whether the supplier can make that decision without quietly transferring the cost of its mistakes to the customer.

Count the work that survives review

Consider a hypothetical publication workflow inside a law firm. An assistant reads a public regulatory announcement, identifies the changes and prepares a first draft for a subject specialist. Generating the draft is one part of the job. Finding an unsupported statement, correcting a misread exception and rewriting an overconfident conclusion are part of the same job.

A cheaper model that produces a usable first draft may be an excellent choice. A cheaper model that saves a small amount of compute but adds half an hour of senior review may be a poor one. Equally, paying for the most capable model on every request can waste money on straightforward formatting and classification.

We would measure accepted output alongside the time taken to check it. For that publication workflow, the review record should show which factual claims required correction, whether the source supported the conclusion and how much rewriting the editor did. Fluent prose is a weak substitute for those observations.

This changes what a useful trial looks like. Give competing systems the same representative source material, include awkward cases and have reviewers assess the outputs without knowing which model produced them. Keep preparation and review time in the comparison. A demonstration selected by the supplier cannot tell a buyer what happens on the cases that fail.

A subscription needs room for success

Usage is an unusual test of an AI subscription. A customer who finds a valuable recurring task may ask the system to do substantially more work while paying the same monthly fee. The supplier then has to absorb the additional cost, improve efficiency or change the commercial terms.

The accompanying chart translates the reported margin percentages into costs for a hypothetical $100 of revenue. It illustrates the arithmetic, not Harvey's actual dollar accounts. Gross margin also differs from operating profit: it does not capture every expense of running the company.

For a buyer, this suggests a practical question before rollout: what changes in the contract if usage rises sharply? A pilot price is much less informative when the eventual arrangement includes limits, overages or different service levels. The answer should be clear before the team builds a routine around the tool.

We would also ask for notice when a material model change affects an established workflow. A successful evaluation is tied to the system that was tested. If that system changes, rerunning a small set of difficult reference tasks is a reasonable condition of continued trust.

The decision belongs at the task level

Harvey's own work points towards a market in which specialised models compete with general models on particular jobs. That gives customers more options, but also makes a single brand name a less complete description of what they are buying.

The strongest suppliers should be comfortable explaining which tasks they can handle cheaply, where they spend more compute and what happens when a result is not good enough. A firm should be able to inspect those decisions without becoming a model laboratory itself.

We would favour a product that demonstrates lower total cost for work its users can accept, and that makes regressions visible when they occur. That leaves plenty of room for specialised models. It also sets a more demanding standard than either a frontier-model badge or a low token price.