On Thursday, OpenAI launched Astra for Law — a product aimed, as Business Insider reported, at the AmLaw 200, the two hundred largest firms in the country. It runs on what the company describes as its most advanced and most expensive model, drawn from the GPT-6 Astra line unveiled September 3, and it ships with legal research plugins and custom model tuning.

Nearly everyone has read this as the opening shot in the war on the billable hour. That reading is wrong, and it is wrong in a way that flatters almost everyone who holds it.

The Billable Hour Was Never the Price

The billable hour is not a cost. It is a format. Clients do not pay for hours the way they pay for electricity; they pay what the matter is worth, and the hours are the paperwork that justifies the number. A firm that shaves four hundred hours off a deal does not send a smaller invoice. It sends the same invoice and books a better realization rate.

Which means the disruption story has the direction of the money backwards. The AmLaw 100 booked $139.7 billion in gross revenue in the 2024 report, up 6.8 percent, with Kirkland & Ellis alone above $7.2 billion. That money is not a meter reading. It is a negotiated price, defended by relationships, risk tolerance, and the durable fact that no general counsel ever got fired for hiring the most expensive firm in the room. Cut the hours per matter and you have not cut the price. You have cut the cost of delivering it. The gap between those two numbers is called profit, and it stays in the partnership.

The Buyer Is the Party With the Pricing Power

Note who the customer is. Not a legal aid clinic, not the solo practitioner above the hardware store, not the forty-lawyer firm in Dayton doing estates and commercial leases. The AmLaw 200 — the institutions that already charge the highest rates in the market and have the least conceivable incentive to pass savings along.

Then note what they are buying. OpenAI built this on the most expensive model it sells. Frontier inference is not yet a commodity, the marginal cost per matter is real, and it is priced for organizations that can absorb it as overhead against a four-figure hourly rate. That is not a quirk of the go-to-market plan. It is the economics. A product whose unit costs only pencil out at the top of the market gets adopted at the top of the market.

So the access-to-justice chorus should be furious, and mostly it is not — because it spent three years telling itself that frontier AI would finally put a lawyer within reach of ordinary people. The pricing says otherwise. The pricing was decided well before the launch event.

OpenAI Keeps the Revenue. The Firm Keeps the Liability.

There is a second thing the disruption framing conceals. As of this month, per a practitioner guide circulating among firm technology committees, there is no public evidence that Astra holds SOC 2, ISO, HIPAA, or comparable compliance certifications. OpenAI’s own Preparedness Framework rates the model Critical for cybersecurity capability, with certain features withheld from users entirely.

Now set that beside the profession it is being sold into. ABA Model Rule 1.6 makes confidentiality the lawyer’s duty, not the vendor’s. Rule 11 makes the signature the lawyer’s. When a filing goes out with an invented citation, or a privileged document lands somewhere it should not, the bar does not sanction the model. It sanctions the person who signed.

This is the part of the arrangement nobody on stage wanted to discuss. The vendor collects the subscription and the model risk stays exactly where it always was — on the firm. In the end, the ethics rules will not be the regulation that governs this technology. The indemnity clause will.

A legal operations director at a four-hundred-lawyer firm, typing from an airport gate between vendor pitches, put it more bluntly than any of the marketing does: “The partners don’t want fewer hours. They want the same fee with fewer people on it. Every deck says efficiency. Every budget conversation is about headcount.”

The Casualty Is the Second-Year Associate

Which brings us to who actually pays. Not the client — the fee survives. Not the equity partner — the spread widens. The people who pay are the ones whose work was always the filler between a partner’s judgment and a client’s signature: the contract attorneys doing document review, the e-discovery vendors, and above all the associate class, whose entire training model rested on doing the tedious thing badly for three years so they might do the hard thing well for thirty.

Firms will keep hiring at the top of the pyramid — in smaller numbers, from a narrower set of schools. The middle of the profession, the tier that taught judgment by grinding, hollows out first.

That may be a fine trade. But it should be argued for honestly, by people willing to say what is being traded. “AI will finally make legal services affordable” is the nicer sentence. It simply is not the one the pricing model wrote.

Sources