The Free Sample: How AI Token Pricing Is Engineered to Feel Cheap

There is a reason the first taste is always the cheap one. The oldest customer acquisition strategy on earth is to make the initial dose feel free, or nearly free, so that the habit forms before the price is fully understood. Gyms sell the first month for a dollar. Streaming services give you thirty days. The corner dealer, in the cruder version everyone recognizes, gives away the sample and charges once the dependency is real. I am not being coy about the analogy, and I am not the first to reach for it, because the structure is genuinely the same: lower the barrier to getting hooked, then monetize the person who can no longer easily walk away. What is new is that the most sophisticated companies in technology have rebuilt this playbook around a unit almost no buyer can intuitively price — the token — and they are using it to make artificial intelligence feel dramatically cheaper than it actually costs.

The clearest recent tell came from Anthropic, and it is worth stating plainly because the move is so elegant it is easy to miss. In April 2026, Anthropic cut the headline price of a Claude Enterprise seat. Depending on the tier, seats that had run anywhere from forty to two hundred dollars a month dropped to a flat twenty dollars. On the procurement spreadsheet, on the line item a CFO actually scans, that reads as a price cut — a generous one. But in the same motion, Anthropic removed the bundled token allowance that used to come with the seat. The seat no longer includes usage. Every token your team spends — in chat, in Claude Code, in the agentic workflows the whole company is now being encouraged to run — is billed on top at standard API rates. The headline number went down. The bill went up. And the part that should make every buyer sit up straight is that on the enterprise usage plan, you cannot disable billing for usage. The meter is not a feature you opt into. It is the floor.

To see the sleight of hand in one picture, put the two pricing models side by side. Under the old deal, a Claude Enterprise seat carried a higher sticker — call it a hundred and twenty dollars in the middle of the old forty-to-two-hundred range — but it bundled a working allowance of tokens, so the number you approved was close to the number you paid, and it was predictable month to month. Under the new deal, the seat is a flat twenty, which reads on the invoice like a two-thirds discount, but the tokens that used to be included are now a separate, uncapped, always-on meter stacked on top. The visible bar shrank. The bar you do not see at approval time is where the money actually went.

The playbook has a name, and it is not new

Before anyone accuses me of melodrama, let me be precise about what is and is not happening, because the manipulation lives in the details and so does the defense. Nothing here is fraud. No vendor is lying about its rates; the API prices are published, the meter is real, the tokens are genuinely consumed. What is happening is a deliberate exploitation of the gap between the price a buyer perceives and the price a buyer pays, and that gap is enormous with AI because almost nobody — not the engineer, not the manager, certainly not the CFO — has an intuitive model of how many tokens a given amount of work consumes. When you cannot price the unit, you anchor on the number you can see. The vendors know this, and they have arranged the numbers you can see to be the flattering ones.

Economists have a clean name for the visible half of this: penetration pricing, or the loss leader. You price the entry point below its true cost to win the account, betting that once the customer is inside, the relationship becomes profitable. That alone is old and mostly benign — supermarkets have sold cheap milk to sell you everything else for a century. What turns a loss leader into a trap is the second ingredient: switching costs high enough that when the price corrects, you cannot leave. And AI has switching costs the milk aisle never dreamed of. Once your engineers have built their muscle memory around one coding agent, once your workflows call one vendor's API, once your prompts and your agent scaffolding and your evaluation suites are all tuned to one model's quirks, moving is not a purchase decision. It is a migration project. The dependency the metaphor points at is not chemical. It is architectural, and it is just as sticky.

The reason I insist on the crude version of the analogy is that the polite versions let everyone off the hook too easily. "Usage-based pricing" sounds neutral, even fair — you pay for what you use, what could be more honest than that? But usage-based pricing is only fair when the buyer can predict usage, and the entire design of the current moment ensures they cannot. The vendor has years of telemetry on how many tokens real workloads burn. The buyer has a slider in a sales deck. That asymmetry is not incidental. It is the product.

Why vendors are selling tokens instead of seats

To defend yourself you have to understand why the industry moved here, and the honest answer is that it moved for two very different sets of reasons — one legitimate and one manipulative — and the vendors are happy to let you believe only the legitimate one.

The legitimate reason is that a seat and a token have fundamentally different cost structures, and for the first time in the history of software that difference actually bites. Classic SaaS had near-zero marginal cost: once Salesforce built the product, one more user logging in cost essentially nothing, so a flat per-seat price worked because the vendor's cost per seat was flat too (roughly zero). AI broke that. Every token an agent generates burns real compute on expensive accelerators, so the vendor's cost is not flat — it scales directly with how hard each user works the model. And the range is staggering. Agentic work is not a single question and a single answer; it is a model calling tools, reading results, and calling again, hundreds of times, for one task. GitHub's own May 2026 research found that agentic coding tasks consume on the order of a thousand times more tokens than a standard single-turn query. A vendor cannot bundle a thousand-x-variable cost into a flat fee without either capping the capability or bleeding money on the heaviest users. So the meter has a real justification, and any honest critique has to concede it.

There is a specific mechanical detail in the Anthropic change that shows how the unbundling doubles as a quiet price increase even before usage climbs. The older enterprise contracts reportedly carried a prepaid token discount in the range of ten to fifteen percent — buy your tokens with the seat and get them cheaper. When the tokens were unbundled, that discount went with them. So a team that keeps its usage exactly flat, changing nothing about how it works, now pays full API rate for tokens it used to get at a discount, on top of a seat that no longer includes them. The headline framing is a price cut. The arithmetic, for anyone whose usage is nontrivial, is a raise. That is the tell that the visible number was chosen for how it reads, not for what it costs.

The manipulative reason sits right on top of the legitimate one, and it is this: a meter lets the vendor advertise a price that is not the price. When cost is bundled into a seat, the buyer sees the whole cost at once and can compare it cleanly to alternatives and to value. When cost is split into a low visible seat plus an invisible meter, the buyer anchors on the low number, approves it easily, and discovers the real total only in arrears, one monthly invoice at a time, after the dependency is already built. The meter transfers forecasting risk from the vendor, who could predict it, to the buyer, who cannot. It captures all the upside of rising usage automatically — as your company leans into AI, which every vendor is aggressively encouraging, your bill climbs with no renegotiation required. And critically, it makes the sticker price a marketing instrument decoupled from the actual cost, which is exactly why Anthropic could "cut prices" and raise revenue in the same announcement.

The receipts: this is an industry pattern, not one company

Anthropic is the cleanest recent example, but singling it out would be unfair and, worse, would let you think switching vendors is the fix. It is not, because the entire industry has converged on the same structure within about a year. Flat-rate AI is being quietly euthanized across the board, and the pattern of how it dies is consistent enough to be predictable.

The most instructive cautionary tale is Cursor, the AI coding tool that was, for a while, the darling of the developer world. In mid-2025 Cursor changed its Pro plan from a request-based model that users understood to a usage-credit model that they did not, wrapping the change in language about "rate limits" that obscured what was really a meter. The result was exactly what the theory predicts: loyal users, already fully dependent on the tool for their daily work, got hit with bills that bore no resemblance to what they thought they were paying. One developer reported three hundred and fifty dollars of overage in a single week. A five-person team burned through forty-six hundred dollars in six weeks — roughly double what they had spent in all of the prior year. The backlash was severe enough that Cursor issued a public apology on the Fourth of July, 2025, and refunded surprise charges. But notice what did not happen: they did not go back to the old model. The meter stayed. The apology was for the communication, not the mechanism, and that distinction is the whole lesson.

The same shape shows up everywhere once you know to look for it. Microsoft made Copilot Cowork generally available in June with a hybrid model that keeps the roughly thirty-dollar seat but adds per-task charges in a new currency it calls Copilot Credits — a meter bolted onto the office suite that never had one. GitHub Copilot, Cursor, and Windsurf have all moved to token-based billing, which is why the industry shorthand is simply that flat-rate AI is dead. The specifics differ, but the direction is unanimous, and unanimity across competitors who agree on nothing else should tell you the incentive is structural, not incidental. When every vendor independently discovers that the meter is better for the vendor, the meter is not a pricing innovation. It is a wealth transfer with good branding.

The bill nobody budgeted for

If this were only a matter of paying a fair price for real compute, it would be a non-story. The reason it is a story is that the meter is producing financial outcomes that organizations did not plan for and cannot control, and the data on this is no longer anecdotal. Zylo's 2026 SaaS Management Index found that seventy-eight percent of IT leaders reported unexpected charges from consumption-based AI pricing. That is not a fringe complaint; it is a supermajority of the people responsible for the budget being surprised by the bill. Enterprise AI spending jumped one hundred and eight percent year over year in 2026, reaching an average of roughly one and a quarter million dollars per organization. And ninety percent of CIOs named AI cost forecasting as their single top deployment challenge — not model quality, not security, not talent, but the inability to predict what the thing will cost.

Read those three numbers together and the picture is unambiguous. Costs are rising fast, the rise is landing as a surprise, and the people who own the budget have flagged forecasting as the thing they cannot do. That is not a market working well. That is a market where the pricing model has deliberately outrun the buyer's ability to reason about it. There is even a grimly funny coda: the same Anthropic that unbundled the tokens also shipped enterprise spend controls to help you manage the bill it just made unpredictable. Selling the fire extinguisher next to the accelerant is not a coincidence; it is the mature phase of the playbook, where the vendor monetizes both the dependency and the anxiety about the dependency. The New Stack captured the mood when it described the whole episode as a tokenomics reckoning. The reckoning is real, and it is being paid by the buyers.

Why "it feels cheaper" is the entire point

I want to isolate the psychological mechanism, because naming it precisely is what makes it defusable. The manipulation is not the meter itself — a meter can be honest. The manipulation is the deliberate construction of a price signal that makes AI feel cheaper than it is at the exact moment a decision is made, and then reveals the true cost only after the decision cannot be reversed.

It works through three well-understood biases, all of them stacked in the vendor's favor. The first is anchoring: presented with a low seat price and an invisible meter, the buyer anchors on the seat and treats the meter as a rounding error, because they have no basis to estimate it. The second is the pain-of-payment gap: a flat fee is felt once, sharply, at purchase, whereas metered cost is diffused across a thousand tiny invisible events — a token here, a tool call there — none of which register as spending in the moment, so the total accumulates below the threshold of notice until the invoice arrives. The third is present bias: the free sample delivers its value now, while the cost arrives later and someone else's budget — next quarter, a different cost center — absorbs it. Every one of these is a documented feature of human decision-making, and the consumption model is engineered, whether by intention or by the ruthless natural selection of what converts, to exploit all three at once.

This is why switching vendors does not save you and why waiting for one honest provider is a losing bet. The pressure that produced this is not the character of any one company; it is the combination of genuinely variable costs and a buyer who cannot price the unit, and that combination rewards whoever makes the cheapest number the most visible. A vendor who priced honestly — one flat number that included realistic usage — would show a higher sticker than its meter-hiding competitor and would lose the deal on the line-item comparison every time. The market is actively selecting against the honest presentation. That is the strongest possible reason the fix cannot be "find the good vendor." The fix has to be on your side of the table.

The defense: treat tokens like the utility they are

Here is the turn, because doom without a remedy is just complaint, and leaders reading this do not need to be told the water is rising — they need to know where the sandbags are. The good news is that this exact problem has been solved before, in a different domain, and the solution transfers almost directly. Cloud computing went through this precise transition fifteen years ago: a shift from predictable owned servers to a metered utility that produced spectacular surprise bills, and the industry answered with a discipline called FinOps — treating variable infrastructure cost as something you actively govern rather than passively receive. AI tokens are the new cloud, and the same discipline applies. The organizations that will not get hurt are the ones that stop treating tokens as an invisible byproduct and start treating them as a metered utility with a budget, an owner, and controls.

The contract-time items reward more thought than they usually get, because this is the one moment the buyer still holds leverage — before the dependency is built, while the vendor still wants the logo. Committed-use discounts are worth chasing, but they carry their own trap that is worth naming: a committed token cap that you overshoot bills the overage at full rate, and one you undershoot is money you prepaid and forfeited, so the commitment only helps if it is sized from real measured usage rather than a sales rep's optimistic projection. This is precisely why visibility has to come before negotiation — you cannot size a commitment you have never measured. The other clause that matters more than its length suggests is portability: the right to export your prompts, your fine-tunes, your evaluation suites, and your workflow definitions. A vendor who makes leaving easy is telling you they intend to keep you by being good. A vendor who makes leaving hard is telling you they intend to keep you whether they are good or not, and the price correction is coming.

Several of these deserve a sentence more, because the difference between knowing them and doing them is where the money leaks. Visibility is first for a reason: the seventy-eight percent who got surprised were, almost by definition, not watching in real time, and the single highest-return move any leader can make this quarter is to instrument token spend so it is as visible as headcount. Model-tier routing is the biggest lever on the actual number, because the habit the industry politely calls "tokenmaxxing" — reflexively reaching for the most expensive model for every task — is where a huge fraction of the waste lives, and it is exactly the waste that the efficiency turn now repricing enterprise AI around output is forcing organizations to confront. And the gateway point is not a throwaway: the same agent gateway that is becoming the control plane for enterprise AI is the natural place to enforce a token budget, because it is the one component every agent action already passes through, which means it is the only place a hard spend cap can actually be applied rather than merely hoped for.

The chart above is the argument for governance in a single frame. In an unmanaged program, a large share of the token bill — the tokenmaxxing, the runaway agents, the duplicated work — is not productive output at all. It is waste that a control plane and a routing policy would eliminate, which means the first and best response to rising AI costs is not to negotiate a better rate or switch vendors. It is to stop paying full frontier prices for work that did not need them and to stop letting unwatched agents burn tokens into the void. You can often cut the bill by a third without touching a single thing your people actually need, and that is the number to chase before you ever pick up the phone to procurement.

What "the right amount of tokens" actually means

The question the smartest leaders are asking is the right one: how do I make sure my people have enough tokens to do their jobs without handing the vendor a blank check? The framing itself is the answer. Tokens are not a perk to be rationed into scarcity, and they are not an unlimited resource to be waved through. They are a measured input to knowledge work, like cloud compute or, frankly, like electricity in a factory, and the right amount is the amount the work demonstrably requires — no more, and crucially no less.

Erring toward scarcity is its own expensive mistake, and it is the one anxious organizations make in reaction to the surprise bills. If you throttle your engineers to control cost, you have simply moved the waste from the invoice to the payroll: a well-paid developer who spends an hour fighting a token cap to save three dollars of model spend is a catastrophic trade, and a team that cannot reach for the capable model when the capable model is what the task needs will quietly lose to a competitor that can. The discipline is not spending less. It is spending deliberately — measuring what each role actually consumes when it is working well, funding that fully, and cutting only the waste that governance reveals rather than the capability the work requires. Get visibility first, and the right budget reveals itself; the number you were afraid to guess turns out to be something you can simply read off the meter you finally bothered to install.

The honest close

Let me end where I started, with the analogy, and be fair to it. The AI vendors are not criminals, and tokens are not narcotics; the comparison is a lens, not an indictment, and the lens is useful precisely because the structure it reveals is real. The first taste is engineered to feel cheap. The dependency, once built, is architectural and hard to escape. And the true price arrives on a meter you were never given the tools to forecast, at a moment when leaving would cost more than staying. Whether any individual product manager intended that sequence is beside the point, because the market is selecting for it regardless of intent — the vendor who hides the meter best wins the deal, so the meter gets hidden.

The thing about knowing you are the mark is that it is the moment you stop being one. None of these tactics survive a buyer who prices the unit, watches the meter, routes work to the right tier, funds real budgets from real measurement, and refuses to sign a contract that will not let them cap spend or walk away. The vendors moved the game to a field where they had all the information and you had a sticker price. The entire defense is to refuse to play on that field — to bring your own instrumentation, your own budgets, and your own governance, and to treat every flattering low number as the beginning of a question rather than the end of one. AI is going to be one of the largest line items your organization has ever carried. The companies that thrive will not be the ones who found the cheapest vendor. They will be the ones who refused to be told what things cost and went and measured it themselves.

Further Reading

• The end of the per-seat model in agentic coding — the structural mechanics of the seat-to-meter shift this piece critiques
• The efficiency turn repricing enterprise AI around output — why tokenmaxxing is the waste to attack first, before you renegotiate a rate
• The agent gateway becoming the control plane for enterprise AI — the enforcement point where a real, hard token budget can actually be applied
• When AI procurement collapsed into universal credits — the broader shift toward metered, credit-based buying this fits inside
• My prediction on where the seat-versus-meter fight lands by 2027 — a falsifiable, dated call on the backlash and the return of predictable pricing