Tokenomics: Why the Cost of Legal AI Is Quietly Rising.

There are two forces pushing the bill up simultaneously. The first is consumption: agentic workflows — where the model plans, drafts, checks and revises across multiple steps — consume orders of magnitude more tokens than a single chat exchange. In one illustrative example from the legal-tech site law.co, a straightforward contract review that took 2,000 tokens as a single prompt ballooned past 200,000 tokens when run as an agentic pipeline. The second force is price: model providers have been raising the cost of their most capable models. The two curves are rising together.

The practical consequence is that a firm’s AI bill in 2026 looks very different from its AI bill in 2024 — not because it has more licences, but because it is running more capable tasks on more capable models, and each of those tasks is burning more tokens than anyone modelled when the contracts were signed.

The traditional legal software pricing model — a flat fee per user per year — was designed for tools where usage is roughly uniform. A lawyer either has access to a research database or they don’t; whether they use it heavily or lightly doesn’t affect the vendor’s costs much. Tokens are different. A lawyer running agentic workflows all day costs the vendor twenty times more to serve than a colleague who logs in once a week to check a definition.

Vendors are adapting. Crosby, an AI-native law firm backed by Sequoia and Cooley, has dropped both the per-seat licence and the billable hour, charging per contract instead — pricing the work done, not the seats filled. The direction of travel is clear: consumption-based pricing is coming to legal AI, and the firms that haven’t modelled what that means for their cost base are in for a surprise.

The shift has a second-order effect on procurement. Legal teams negotiated their current AI contracts when usage was light and tasks were simple. Renewal conversations are now happening against a very different usage profile, and the leverage has shifted. A firm that hasn’t tracked token consumption across its tools doesn’t know what it’s actually spending — or what it should be paying.

The vendors building on top of foundation models are not sitting still. Three cost-reduction strategies are emerging, each with different implications for buyers.

Precise pricing data for the Australian legal AI market is hard to come by — vendors quote on application, contracts are confidential, and list prices are rarely the prices anyone pays. The following table reflects what is publicly known or widely reported as of mid-2026. It is directionally accurate, not a procurement guide.

Substantial discounts off opening prices are widely reported, especially for multi-year commitments and larger seat counts. No vendor publishes Australian pricing publicly; figures are converted from reported US rates and should be confirmed with vendors before any procurement decision.

Thomson Reuters’ 2025 survey found a majority of active law-firm users now use generative AI at least weekly, and roughly a third daily. But frequency is not intensity: the lawyers running million-token agentic workloads — the ones who actually move the bill — remain a small minority. (Thomson Reuters’ 2026 follow-up already puts weekly use among current users above 80%, so even the frequency baseline is climbing.)

That will change. The useful analogy is a learner driver: a little fuel, on local streets, at low cost. The same driver, years on, takes the car cross-country at speed. Most firms are at the learner stage. As more lawyers become confident, agentic users, consumption — and cost — will rise several-fold on top of any price increases.

The takeaway is not alarm. Large firms can comfortably absorb these costs; the point is that AI spend is additive to a technology stack — document management, research subscriptions, the Microsoft suite — that firms cannot drop. It swells the total budget beyond where it sits today. The firms that fare best will be the ones that model that curve now, rather than discovering it later.