Token Economics in Legal AI

Picture your firm choosing a legal AI tool. Two vendors give great demos. Both review contracts, pull up case law, and draft in seconds. You compare the annual prices, pick the one that fits your budget, and sign.

Six months later, the bill looks nothing like what you expected.

Nobody misled you. You just looked at the wrong number. Underneath the subscription price of almost every legal AI tool, there is a second layer: a usage meter that decides what the tool really costs once your whole team starts using it. Most buyers never look at it. This piece is about why you should.

What A Token Actually Is

Every legal AI tool runs on a large language model, the engine inside products like Harvey, CoCounsel, or an in-house contract assistant. These models read and write in small units called tokens.

A token is not a word. It is a piece of one, about four characters on average, or roughly three-quarters of a word. A 1,000-word document comes to around 1,300 to 1,500 tokens. The word "lawyer" might be one token, while "indemnification" might be three or four.

The companies that build these models charge by the token, and here is the part that matters. They charge twice: once for everything you put in (your question, the contract, the case file), and again, at a higher rate, for everything the model writes back (the summary, the redline, the advice).

Output usually costs around five times more than input. The reason is simple: reading is easy, writing is hard. Generating new text takes the model far more computing power than reading text you already gave it, the same way a lawyer can skim a contract quickly but takes much longer to draft one.

The actual rates are public. As of mid-2026, a normal working model like Claude Sonnet 4.6 or GPT-5.4 costs around $2.50 to $3 per million input tokens and about $15 per million output tokens. Budget models cost a fraction of that. The most powerful reasoning models, which think in steps before answering, can reach $30 input and $180 output per million, and since you pay for each thinking step too, the real cost runs higher.

Why Legal Work Runs Up The Bill

Legal work is unusually expensive to run through AI for two reasons every lawyer will recognize. First, the documents are long. A commercial contract, a bundle of authorities, and a due diligence data room,these are huge inputs. Feed a 100-page agreement into a model and ask it to check every clause, and you have spent a serious number of tokens before it writes a word back. Second, the answers are long too. A real clause-by-clause review, or a proper redline, is a big output, and output is the expensive side.

Think about the contrast. Someone asking a chatbot for a recipe spends a trickle of tokens. A lawyer asking AI to review a 90-page agreement against a playbook spends a flood. So in a legal setting, pricing deserves more attention, not less.

This is also why legal AI does not behave like normal software. Ordinary software costs almost nothing to give to one more user: build it once, and the millionth user is nearly free. AI is the opposite. Every query costs the vendor real money, because every query burns computing power. The more your lawyers use the tool, the more it costs to run.

Will Rising Costs Break The Model?

There is a fair worry here. If legal teams start using AI everywhere, won't these costs spiral until the whole thing becomes unaffordable? The evidence points the other way, with one catch.

Per token, prices have dropped sharply. A job that cost about as much as a dinner three years ago now costs less than a cup of coffee. Four big AI labs are in a price war, and cheaper open models keep pushing rates down. Token for token, legal AI is getting cheaper.

The catch is how much gets used. As the tools improve, handling longer documents, taking more steps, and reasoning before answering, each task quietly uses far more tokens than before. Falling prices meet rising use, so the bill does not necessarily shrink. It just gets harder to predict. The risk is no longer that one query is expensive. It is that thousands of queries, run by people who never see the meter, add up to a number nobody planned for.

Should Firms Pass These Costs to Clients?

This is where the numbers meet how firms actually bill, and there is no settled answer yet.

The traditional model bills time, and AI saves time: a research task that took a junior a full day might now take an hour. If you bill by the hour, AI seems to shrink your own invoice. If you treat it as a cost of doing business, you absorb the bill and compete on speed and price. Some firms will try to recover the cost as a disbursement, the way they once charged for paid databases. But clients are getting more AI-aware, and many will push back on paying extra for what looks like the firm's own efficiency tool. Others will fold AI into fixed-fee deals. Kirkland's own chair told the Financial Times that AI would speed up exactly this shift away from the billable hour.

The clearest signal comes from the top of the market. In May 2026, Kirkland & Ellis, the highest-earning law firm in the world, said it would spend around $500 million over three to four years building its own AI tools instead of renting them. The platform draws on input from 250 of its own lawyers, and the outside companies helping build it cannot resell it. Owning the technology means owning the cost, the data, and the roadmap.

For Indian firms, the lesson from Kirkland is not the one it first looks like. Kirkland can spend $500 million because it earns over $10 billion a year, and no Indian law firm comes anywhere near that scale, so copying the build-your-own path is not realistic here. The real takeaway runs the other way. Since Indian firms will keep renting AI rather than building it, the smart move is to be careful about which vendors they choose and how those tools are priced. Indian firms are also far more cost-sensitive than the global giants, and the unpredictable credit-based pricing that many Indian legal AI tools now use can hurt a smaller practice much faster than a large one. For an Indian firm, control over cost is not a luxury. It is the whole game.

The Questions A Buyer Should Ask

For anyone choosing legal AI, whether a managing partner, a general counsel, or a legal operations lead, this comes down to a few practical questions that matter more than the demo.

How is the tool priced, a flat subscription or credits that run down with use? A flat fee puts the risk on the vendor. A credit model puts it on you. Neither is wrong, but they behave very differently once a whole team uses the tool every day. Indian legal research providers have increasingly moved to this split, a base subscription plus AI features billed on credits. It is honest, because the AI really does cost them money per query, but it is unpredictable, because a junior still learning to use it can burn through a month's credits in a week.

Can you see and cap what you spend? A tool with no per-user tracking is one whose real cost you only find out at renewal. Ask whether usage can be capped per person and reviewed each month. If it can't, the budget is not really in your control.

What is the tool doing under the hood? A product that quietly sends every routine query to the most expensive model will cost far more than one that saves the premium engine for hard problems and routes the rest to cheaper models. That one choice can move the bill several times over.

Final Thoughts

The honest way to judge a legal AI tool is not its monthly price but what it costs to get real work done well, at the quality the matter needs. That includes what the sticker price hides: use that grows as more people adopt it, the premium for reasoning-heavy models, and the time spent checking the output, because an unchecked answer in law is worth less than nothing.

Legal AI is no longer really about whether the technology works. It does. The open questions now are commercial, and cost sits right at the center of them. The tools that last will not always be the most powerful. They will be the ones whose cost makes sense at scale: easy to budget for, easy to track, and priced in a way that survives a whole office using them every day.

For everyone choosing legal AI, the lesson is simple. Look past what the tool can do, and ask what it will cost to keep doing it.

---

This article has been authored by Aditya Pratap Singh, LegalTech Fellow at the Indian LegalTech Network and a student at Llyod Law College.