There has been a lot of buzz around AI token pricing lately, especially as agents create a multiplier effect, using an unpredictably large number of AI resources known as tokens.
Uber burned through its 2026 AI budget as it rolled out Claude Code to their developers. The term Token Maxxing surfaced in the software engineering community, where coders were measured on how much AI they used in their job without regard to efficiency or thought. Tokens have been subsidized by foundation model businesses in some instances to get users hooked, but that is ending.
The legal industry is beginning to feel some of this uncertainty, too. Legora recently moved away from a per-seat model and now charges for its Agent Pro solution based on consumption and resource intensity. This pricing model follows the lead of larger players like Anthropic and shifts the cost risk onto users of Legal AI. That creates issues with budgeting, forecasting, and using AI solutions efficiently.
The concerns are understandable. As models become more capable, users and agents process more information, submit larger prompts, upload more documents, and run increasingly sophisticated workflows. Every improvement seems to encourage more consumption.
But before we assume this is an entirely new problem, it is worth remembering that the legal industry has already lived through something remarkably similar on more than one occasion.
The legal community has seen this movie before.
Comparisons have been made to legal research pass-through charges to clients. I want to go way back to the Lexis Search Unit.
Law Firms Hated the Lexis “Search Unit”
In the 1980s, online legal research was transforming the practice of law. The original Lexis service was expensive, but it delivered enormous value by surfacing cases that would otherwise have been missed when researching issues in print. The core pricing was time-based with some ancillary charges. For example, printing cases cost 2 cents per line. But the fundamental pricing model was relatively easy to understand. Lawyers knew that spending more time online would cost more money.
Then Lexis incorporated an additional variable in its pricing model called Search Units.
Search Units were very similar to AI tokens. They reflected the intensity of the computing resources consumed during a search. From the Lexis view, the model reflected their cost structure. But how could a user understand the resources their search would take before submitting it? The model transferred risk to users without offering ways to manage it. This was maddening, and the legal community hated it.
Sound familiar? How does a law firm know how many tokens an AI agent will consume?
There was another subtle component to the Search Unit model. As more documents were added to the database, searches became more computationally intensive. Lexis added new documents every day, and when the same search was run a few weeks later, it would cost more Search Units. There was a built-in inflation factor.
The same is likely to be true with token-based pricing. The more advanced the LLM, the more tokens it consumes, and the same applies to more powerful agents.
Law firms saw their bills rise under Search Units, and they should expect the same with token-based pricing.
The coming demand for predictability
History often rhymes and even repeats itself. Rising consumption will create a backlash. Assuming there is value in AI, the backlash won’t be about cost. It will be about the uncertainty of budgeting and planning. This is especially true when the investment community is focused on the benefits of AI initiatives.
Usage-based pricing based on tokens will hit a breaking point. Before we explore that, let’s consider another question.
Could tokens become the new billable hour for agentic work?
The billable hour has been the proxy for value in the legal profession for decades. On the surface, AI agents based on consumption models parallel the billable hour.
If clients can think of consuming a lawyer’s time by the hour, what is wrong with paying for an AI agent that approximates the work of a staff member? Could firms assign an hourly rate for AI agents or pass through token costs with a modest markup?
Perhaps. Law firms can estimate a lawyer’s billable hours and set expectations for a client. But the inability to estimate the cost of an AI agent creates challenges. Changing outside counsel guidelines isn’t always a simple task either. If law firms can offer predictability for their AI agents, perhaps they can convince clients to pay for their direct costs.
A more likely possibility is a change in vendor pricing models
The Lexis Search Unit has many parallels to tokens. So, how did Lexis remedy the Search Unit problem?
In the early 1990s, Lexis began offering new packages with clearer, more predictable pricing. They modeled usage and proposed pricing based on historical growth patterns. The hook was a discount based on a commitment to continue growing usage. These “commitment pricing” packages were among the first multi-year agreements for legal research, and they helped normalize custom pricing that varies widely from firm to firm. Most of all, it solved the budgeting issue by offering predictability.
I think it’s reasonable to assume that once enterprise usage patterns are established, token-based pricing will be replaced by predictable models that enable enterprises to budget.
The Lexis commitment pricing packages illustrate how this could evolve. Lexis offered other sweeteners that added value, and I think vendor negotiations for AI solutions could do the same.
The hidden benefit for vendors
Predictability in pricing benefits customers. Not only do vendors benefit from predictable revenues, which provide investors with comfort, but they also flip the incentive from “maxxing” consumption to driving efficiency. Once revenue is locked in, they are motivated to save, creating an opportunity to reduce costs and improve margins.
As Lexis locked in revenues with commitment pricing, it also benefited from Moore’s Law as computing costs fell.
AI vendors will have an incentive to deploy specialized, token-efficient models for specific tasks. To the extent there is a Moore’s Law for AI, they’ll benefit from that. Deep Research might think less before providing answers, and AI agents may come with built-in throttles and guardrails to make them more efficient or prevent a digital “fool’s errand.”
Token consumption is a remake of the Search Unit movie
Vendors, starting with OpenAI and Anthropic, need to grow revenue to stop losing money. Charging for tokens is one way they can do that in the short term.
Lawyers did not buy legal research because they loved Search Units. They bought legal research because they needed answers. Similarly, lawyers will reluctantly pay for tokens to achieve better results with AI solutions.
History rhymes. This is a transitional pricing model. There will be a sequel to the token consumption movie.
Budget predictability is the problem to be solved with token-based pricing. Vendors will eventually have to fix this for customers while maintaining a fair margin. Over the long term, successful vendors won’t charge for tokens. They’ll be the ones who get customers to think about value.
AI was used in the creation of this article
Ken Crutchfield has over forty years of experience in legal, tax, and other industries. Throughout his career, he has focused on growth, innovation, and business transformation. His consulting practice advises investors, legal tech startups and others. As a strategic thinker who understands markets and creating products to meet customer needs, he has worked in start-ups and large enterprises. He has served in General Management capacities in six businesses. Ken has a pulse on the trends affecting the market. Whether it was the Internet in the 1980s or Generative AI, he understands technology and how it can impact business. Crutchfield started his career as an intern with LexisNexis and has worked at Thomson Reuters, Bloomberg, Dun & Bradstreet, and Wolters Kluwer. Ken has an MBA and holds a B.S. in Electrical Engineering from The Ohio State University.
The post Token Maxxing: This Pricing Movie Is A Remake appeared first on Above the Law.
There has been a lot of buzz around AI token pricing lately, especially as agents create a multiplier effect, using an unpredictably large number of AI resources known as tokens.
Uber burned through its 2026 AI budget as it rolled out Claude Code to their developers. The term Token Maxxing surfaced in the software engineering community, where coders were measured on how much AI they used in their job without regard to efficiency or thought. Tokens have been subsidized by foundation model businesses in some instances to get users hooked, but that is ending.
The legal industry is beginning to feel some of this uncertainty, too. Legora recently moved away from a per-seat model and now charges for its Agent Pro solution based on consumption and resource intensity. This pricing model follows the lead of larger players like Anthropic and shifts the cost risk onto users of Legal AI. That creates issues with budgeting, forecasting, and using AI solutions efficiently.
The concerns are understandable. As models become more capable, users and agents process more information, submit larger prompts, upload more documents, and run increasingly sophisticated workflows. Every improvement seems to encourage more consumption.
But before we assume this is an entirely new problem, it is worth remembering that the legal industry has already lived through something remarkably similar on more than one occasion.
The legal community has seen this movie before.
Comparisons have been made to legal research pass-through charges to clients. I want to go way back to the Lexis Search Unit.
Law Firms Hated the Lexis “Search Unit”
In the 1980s, online legal research was transforming the practice of law. The original Lexis service was expensive, but it delivered enormous value by surfacing cases that would otherwise have been missed when researching issues in print. The core pricing was time-based with some ancillary charges. For example, printing cases cost 2 cents per line. But the fundamental pricing model was relatively easy to understand. Lawyers knew that spending more time online would cost more money.
Then Lexis incorporated an additional variable in its pricing model called Search Units.
Search Units were very similar to AI tokens. They reflected the intensity of the computing resources consumed during a search. From the Lexis view, the model reflected their cost structure. But how could a user understand the resources their search would take before submitting it? The model transferred risk to users without offering ways to manage it. This was maddening, and the legal community hated it.
Sound familiar? How does a law firm know how many tokens an AI agent will consume?
There was another subtle component to the Search Unit model. As more documents were added to the database, searches became more computationally intensive. Lexis added new documents every day, and when the same search was run a few weeks later, it would cost more Search Units. There was a built-in inflation factor.
The same is likely to be true with token-based pricing. The more advanced the LLM, the more tokens it consumes, and the same applies to more powerful agents.
Law firms saw their bills rise under Search Units, and they should expect the same with token-based pricing.
The coming demand for predictability
History often rhymes and even repeats itself. Rising consumption will create a backlash. Assuming there is value in AI, the backlash won’t be about cost. It will be about the uncertainty of budgeting and planning. This is especially true when the investment community is focused on the benefits of AI initiatives.
Usage-based pricing based on tokens will hit a breaking point. Before we explore that, let’s consider another question.
Could tokens become the new billable hour for agentic work?
The billable hour has been the proxy for value in the legal profession for decades. On the surface, AI agents based on consumption models parallel the billable hour.
If clients can think of consuming a lawyer’s time by the hour, what is wrong with paying for an AI agent that approximates the work of a staff member? Could firms assign an hourly rate for AI agents or pass through token costs with a modest markup?
Perhaps. Law firms can estimate a lawyer’s billable hours and set expectations for a client. But the inability to estimate the cost of an AI agent creates challenges. Changing outside counsel guidelines isn’t always a simple task either. If law firms can offer predictability for their AI agents, perhaps they can convince clients to pay for their direct costs.
A more likely possibility is a change in vendor pricing models
The Lexis Search Unit has many parallels to tokens. So, how did Lexis remedy the Search Unit problem?
In the early 1990s, Lexis began offering new packages with clearer, more predictable pricing. They modeled usage and proposed pricing based on historical growth patterns. The hook was a discount based on a commitment to continue growing usage. These “commitment pricing” packages were among the first multi-year agreements for legal research, and they helped normalize custom pricing that varies widely from firm to firm. Most of all, it solved the budgeting issue by offering predictability.
I think it’s reasonable to assume that once enterprise usage patterns are established, token-based pricing will be replaced by predictable models that enable enterprises to budget.
The Lexis commitment pricing packages illustrate how this could evolve. Lexis offered other sweeteners that added value, and I think vendor negotiations for AI solutions could do the same.
The hidden benefit for vendors
Predictability in pricing benefits customers. Not only do vendors benefit from predictable revenues, which provide investors with comfort, but they also flip the incentive from “maxxing” consumption to driving efficiency. Once revenue is locked in, they are motivated to save, creating an opportunity to reduce costs and improve margins.
As Lexis locked in revenues with commitment pricing, it also benefited from Moore’s Law as computing costs fell.
AI vendors will have an incentive to deploy specialized, token-efficient models for specific tasks. To the extent there is a Moore’s Law for AI, they’ll benefit from that. Deep Research might think less before providing answers, and AI agents may come with built-in throttles and guardrails to make them more efficient or prevent a digital “fool’s errand.”
Token consumption is a remake of the Search Unit movie
Vendors, starting with OpenAI and Anthropic, need to grow revenue to stop losing money. Charging for tokens is one way they can do that in the short term.
Lawyers did not buy legal research because they loved Search Units. They bought legal research because they needed answers. Similarly, lawyers will reluctantly pay for tokens to achieve better results with AI solutions.
History rhymes. This is a transitional pricing model. There will be a sequel to the token consumption movie.
Budget predictability is the problem to be solved with token-based pricing. Vendors will eventually have to fix this for customers while maintaining a fair margin. Over the long term, successful vendors won’t charge for tokens. They’ll be the ones who get customers to think about value.
AI was used in the creation of this article
Ken Crutchfield has over forty years of experience in legal, tax, and other industries. Throughout his career, he has focused on growth, innovation, and business transformation. His consulting practice advises investors, legal tech startups and others. As a strategic thinker who understands markets and creating products to meet customer needs, he has worked in start-ups and large enterprises. He has served in General Management capacities in six businesses. Ken has a pulse on the trends affecting the market. Whether it was the Internet in the 1980s or Generative AI, he understands technology and how it can impact business. Crutchfield started his career as an intern with LexisNexis and has worked at Thomson Reuters, Bloomberg, Dun & Bradstreet, and Wolters Kluwer. Ken has an MBA and holds a B.S. in Electrical Engineering from The Ohio State University.
