Tokenomics Explained: What AI Actually Costs, and Why Your Use Cases Are the Only Honest Answer

I've always been fascinated by prices. Not the numbers themselves, but what they reveal. A good price tells you the truth about what something costs to make and what it's worth to the person buying it. A bad price hides one of those two things, and sooner or later the hidden thing comes looking for you.

Right now, AI pricing is hiding both.

‍

If you lead a pharmaceutical company, a research organization, or a university, you're being asked to make real budget commitments to AI at the exact moment its pricing is least honest. Flat monthly subscriptions sit next to metered APIs, which sit next to "AI included" line items buried in SaaS renewals. The first question that comes to mind is: "What does this actually cost us, and what are we actually getting for it?" Answering that question requires understanding something the FinOps community has started calling tokenomics.

The atomic unit of AI value

Every interaction with a large language model is measured in tokens, small fragments of words, roughly 2,000 tokens for every 1,500 words of English by OpenAI's own rule of thumb. Tokens are the atomic unit of AI. Every prompt you send consumes them, every response you receive is made of them, and every one of them costs real compute, which costs real electricity and which costs real money.

‍

This matters because tokens are where AI pricing stops being abstract. You can argue about what a "seat" of AI software is worth. You cannot argue about how many tokens a workload consumed. Tokenomics is simply the discipline of connecting the tokens you consume to the business outcomes they produce. A workflow that consumes ten times the tokens but produces a hundred times the value is a bargain. The goal is to know the ratio.

‍

Here's the part that surprises a lot of people: token consumption is wildly nonlinear. A simple question posed directly to a small model might consume a few hundred tokens. That same question routed through an agent, one that searches your document repository, reasons through the results, calls a few tools, retries a failed step, and drafts a response, can consume a hundred times more. The FinOps Foundation's analysis of token economics puts agentic and reasoning workloads at five to thirty times the token consumption of simple chat, and identifies them as the principal driver of enterprise AI spend growth today.

Anyone who works in a lab already understands this way of thinking. Research organizations have in recent history lived on unit economics (cost per assay, cost per sample, cost per sequenced genome). Nobody runs a lab on an "unlimited reagents" plan, because reagents have marginal costs and everybody knows it. Tokens follow similar realities. The organizations that internalize this early will make dramatically better AI decisions than the ones that don't.

‍

Why the flat fee can't survive agents

Which brings us to the $20 problem.

‍

The flat AI subscription was never a price; it was an anchor. As Rishi Sidhu put it in his piece on why the $20 AI subscription is dying, it was the milk at the back of the store, set below cost to pull the market in during the land grab phase. That worked fine when AI meant a person typing questions into a chat box, because human attention is a natural rate limiter. You can only type so fast.

‍

Agents removed the rate limiter. An agent working on your behalf can run all night, fan out into dozens of parallel tasks, and burn compute at a pace no human ever could. And unlike streaming video, where one more viewer costs the provider almost nothing, every one of those tokens carries a direct computational cost. This inverts the entire logic of subscription software. Your heaviest users, the ones getting the most value, the ones you built the product for, become your least profitable customers.

We're already watching the timer run out. Frontier model providers have publicly acknowledged that unlimited plans don't work and are moving to structures that combine seats with committed consumption. AI coding tools have shifted from flat tiers to metered billing, and their customers watched effective spend jump by an order of magnitude. The pattern is consistent, and I'd state it as close to a law as anything in this industry. Every flat-fee, agentic product will transition to tokenomics, and this will remain true until the cost of compute becomes astronomically low. That day may come. The International Energy Agency reports that AI data center electricity demand grew 50% last year against 3% global electricity growth, and demand is compounding faster than efficiency gains.

‍

So when a vendor offers your organization unlimited agentic AI for a flat fee, they are telling you one of three things: They're subsidizing you with investor money that will run out. They're quietly throttling you in ways you can't see. Or they've priced in a margin so large that you're subsidizing their other customers. None of these is a foundation for a partnership meant to last years. In industries like pharma and higher education, where budgets are set in annual cycles, and a surprise repricing can't be absorbed in the middle of a fiscal year, that foundation matters more than almost anywhere else.

‍

Start with the use case, not the price tag.

It's easy to read about tokenomics and conclude that the job is to hunt for the cheapest tokens. That is not the goal. A token has no value at all until it's attached to a use case, and a use case has no value until you can name what it's worth to your organization.

‍

Let's say you work at a research university and you're evaluating an AI agent that helps labs source and purchase supplies. The token cost of each procurement interaction is knowable. The real question is the other side of the equation. What does it cost you today when a postdoc spends four hours hunting for a reagent across supplier catalogs? What does a procurement delay of three weeks cost a study running on grant money with a fixed end date? What's the value of catching a 30% price variance on an antibody your labs buy four hundred times a year? When you can put even rough numbers on those questions, the economics stop being mysterious. Cost per outcome, whether that outcome is a sourced order, a screened compound, or a drafted protocol, becomes something you can actually evaluate, compare, and defend to a CFO or a grants office.

‍

Organizations that skip this step end up in one of two failure modes. They either dramatically overpay for AI that produces activity rather than outcomes, or they starve the workloads where the return is largest: deep research and complex agentic workflows that consume a lot of tokens and earn every one of them. Only a small fraction of organizations can attribute meaningful earnings impact to AI today, and I'd argue the difference isn't the technology they bought. It's that the high performers did the unglamorous work of mapping use cases to value before they signed anything.

How we've attacked this at Labviva

I'll tell you how we've approached this ourselves, because I think it illustrates what's possible when a company treats cost as an engineering problem rather than a pricing problem.

‍

At Labviva, we made a decision early on that we would not simply route every task in our agentic toolset to the largest frontier model and pass the bill along. Instead, we've been deliberate about matching each unit of work to the smallest, cheapest model that can do it well. A frontier reasoning model is remarkable, and some tasks in scientific procurement genuinely need one. Most do not need this: classifying a product, matching a catalog item, extracting terms from a quote. These are jobs for small language models built for exactly that purpose, and using a reasoning model to do them is like using a mass spectrometer to determine the mass of your lunch.

‍

That's only possible because of the second decision we made: to invest in our own internal data capabilities and in models built specifically for our customers' domains. Years of structured data about life science products and their use mean our agents don't have to reason their way to answers from scratch, burning tokens all the way, when the knowledge is already sitting in our data, a cheap lookup away. Small language models trained on the language of scientific procurement handle the specialized work at a fraction of the token cost of a general-purpose giant. The result is that the expensive models get reserved for the moments that actually need them and the cost per outcome keeps going down instead of up.

Choose partners, not price tags.

This raises a question: if you're a pharma company, a research organization, or a university, what should you actually look for in an AI provider?

‍

Look for a partner who understands your domain. General-purpose AI burns enormous numbers of tokens rediscovering things a domain expert already knows. A partner with deep roots in your world, who knows the suppliers, the workflows, the compliance realities, and the vocabulary of a lab, has already paid that cost down into data and domain models, and it shows up directly in your unit economics.

‍

Look for a partner who is visibly focused on driving cost out, and who can tell you exactly how. Ask them where they use small language models instead of frontier ones. Ask how they route work between models, what lives in structured data versus what gets reasoned over, and how their cost per outcome has trended. A partner doing this work will light up at the question. A partner who changes the subject is planning to make their economics your problem.

‍

Look for transparency in the unit cost and in how the meter runs. You should know what a unit of work costs, whether that unit is an interaction, a workflow, or an outcome, and you should be able to see and forecast consumption the way you forecast any other operational cost. "Trust us, it's included" is not transparency; it's deferred repricing.

‍

And look for a partner who will sit down and tailor the solution to your processes, behaving like a knowledge consultant rather than a software vendor. The truth is that most organizations don't yet know which of their use cases justify heavy token spend and which don't, and they shouldn't be expected to. A real partner works alongside your scientists, your procurement team, and your administrators, maps the solution to how your organization actually works, and helps you find out. They'll tell you when a cheaper model is good enough. They'll tell you when a workflow isn't worth automating at all. A vendor who only ever recommends more consumption is a vendor whose incentives you now understand perfectly.

‍

The subsidized era of AI pricing was pleasant while it lasted, but its end is good news for serious organizations. Honest prices make honest decisions possible. The token is just the atom. The value was always in what you build with it, and the organizations that know exactly what they're building and exactly what it's worth are the ones for whom AI will be the bargain of the decade.

‍

References: Token Economics: The Atomic Unit of AI Value (FinOps Foundation), Why the $20 AI Subscription Is Dying (Rishi Sidhu), What are tokens and how to count them? (OpenAI), and Data centre electricity use surged in 2025 (International Energy Agency, April 2026).

‍