Our Take
The legal tech market built itself on the assumption that LLM costs would stay cheap and per-seat pricing would work forever. Both assumptions are now broken.
Why it matters
Law firms scaling AI to document-heavy tasks face spiraling costs they can no longer absorb or hide. The economics of legal AI—which shaped every vendor's business model—are shifting in real time, forcing hard choices about which models to use and whether to build in-house.
Do this week
General Counsel: audit your firm's current AI spending by task type this week so you can benchmark against peers and decide whether to move to cheaper models, fine-tune open-source alternatives, or renegotiate vendor contracts before Q3.
Token prices are rising faster than legal AI adoption can absorb
OpenAI and Anthropic have raised prices on their latest frontier models (GPT-5.5, Claude Opus 4.7) at the same time law firms are using them harder. The reason is straightforward: newer models trained on enterprise usage data rather than internet crawls are smarter and more aligned to professional tasks, so lawyers love them. Agents add another layer of cost—they read documents repeatedly, reason through problems, and consume orders of magnitude more tokens than single-pass queries.
The result is a collision between rising demand and rising unit costs. Shawn Curran, CEO of Jylo, told Artificial Lawyer: "Because the internet is a commodity there was competition, now OpenAI and Anthropic have the bulk of users, they have the data fly-wheel, and are pulling ahead and can set the price without too much competition." Per-seat pricing—the business model almost every legal tech vendor sold—is now dead. Vendors can no longer subsidize all-you-can-eat consumption. Law firms will have to spend more, and they will have to choose.
The legal tech margin game is collapsing
For five years, legal tech vendors bought tokens cheap and resold them at a markup, wrapping them in workflow tools and proprietary UI. That arbitrage worked because foundation model prices fell and adoption was still new. Now the margin is gone. OpenAI and Anthropic own the relationship, control the price, and (per Antti Innanen, legal AI expert behind Laverne) "the new models were heavily subsidized at first to attract users and developers. I am not entirely sure what the long-term strategy is for API pricing."
Three responses are emerging. Some vendors are routing expensive tasks to cheaper, older models and reserving frontier models for cases that actually need them. Thomson Reuters and Kirkland & Ellis are building fine-tuned open-source models to reduce dependency on Claude and ChatGPT entirely. Harvey is working with Factory to route work dynamically and with LangChain to reduce verifier costs—bringing down verification overhead by an order of magnitude by batching and using open models instead (company-reported).
The strategic divergence is real. Jake Jones, co-founder of Flank, reframed the math: "Work is shifting to agentic workloads that use orders of magnitude more tokens. Even as models get smarter and cheaper, longer-running tasks will keep consumption climbing. That's already shifted our pricing away from 'this is a tool' towards 'this is an autonomous system displacing humans', priced on the displacement, not the licence." In other words: it's expensive, but you get what you pay for.
Three moves to control legal AI costs now
First, stop assuming your vendor absorbs token inflation. Ask directly how your legal tech supplier is managing cost changes and whether they pass them on. Expect the answer to shift from "we don't" to "we do, starting next quarter."
Second, audit which tasks truly need frontier models. Escalation protocols—routing complex problems to GPT-5.5, routine work to GPT-4 or open models—cut costs without sacrificing quality. Most document review and basic contract tagging do not need the latest model.
Third, watch the fine-tuning moves. If you have the budget and the data, building a legal-specific open-source model (as Kirkland and Thomson Reuters are doing) reduces long-term token dependency. If you don't, pressure your vendor to do it on your behalf or negotiate a fixed token budget instead of per-seat pricing, so at least you know the damage.
The old legal tech economics—cheap tokens, subsidized per-seat deals, vendor margin—is over. The question now is whether you choose how to adapt or wait for your vendor (or OpenAI and Anthropic) to choose for you.