Stop Banking the Discount: Why Cheaper AI Should Make You Spend More, Not Less

In short

• Economics. The blended price of a million tokens fell from about $17 to $2 in a year — roughly 8x — while usage grew ~14x. Cheaper inputs drive more use, not less.
• Surplus. The value firms place on AI runs ~30% above what they pay for it. A discount this large will not last.
• Trap. 7 in 10 corporate AI claims are about cost savings. Efficiency banks a discount; new work compounds.
• Portfolio. “Zero loyalty” is now rational. Frontier models plan and handle the hard parts; cheaper and open models do the bulk of execution.
• Independence. Buy AI like a procurement category — model-agnostic architecture, contract discipline, no single-lab lock-in.

There are two true stories about the cost of AI, and they appear to contradict each other. The first is the one every board heard this year: AI is expensive and getting more so. The second, quieter story is that AI has become dramatically cheaper — the blended price of a million tokens has fallen from roughly $17 to about $2 in a year, an eightfold drop, according to Exponential View’s State of the AI Economy report, summarised in Inc. Both are true. Reconciling them is the most important budgeting decision most companies will make this year — and the evidence suggests most will get it wrong.

When the price of something useful falls, a first-year economics student can tell you what happens next: you do not use the same amount and pocket the difference. You find uses you never bothered with when it was expensive. That is the whole story of AI spend in 2026, and it points to a conclusion that will feel counter-intuitive to every CFO reaching for the savings: cheaper AI is a reason to spend more, not less — provided you spend it on work you could not do before, and buy it like a procurement function rather than a fan.

The price war you’re winning — and why it won’t last

The leading AI labs are, by every account, in a price war for enterprise business — and the customer is the beneficiary. The Wall Street Journal chronicled vendors offering incentives aggressively to lock customers in: one startup, Pylon, estimated it had received around $1.6 million in free tokens from a single vendor this year, with tens of thousands more from others. Exponential View’s report, via Inc., puts a figure on the windfall: drawing on Stanford’s Digital Economy Lab, it found the value people place on AI runs roughly thirty percent above what they actually pay for it. Economists call that gap consumer surplus, and right now it is unusually large. The strategic point is the second half of that sentence: a price war is a temporary condition. You may never get this much value per dollar again — which makes this the moment to buy ambitiously, not frugally.

Jevons in the boardroom: cheaper AI means more spend, not less

In the 19th century the economist William Stanley Jevons observed that making coal use more efficient did not reduce coal consumption — it increased it, because efficiency made coal worth using for more things. AI is running the same play, at speed. Exponential View’s report, per Inc., measures the elasticity directly: every ten percent cut in token prices drives twelve to eighteen percent more usage, so total spend rises even as the unit price falls. The volumes are staggering — one large provider went from processing under 10 trillion tokens a month to 480 trillion, and industry token volumes are growing roughly fourteen times a year. The same report notes AI revenue is compounding about three times faster than the internet or mobile waves did. Falling prices here are not deflation. They are an accelerant.

The efficiency trap: banking a discount vs. compounding new work

Here is where most companies will go wrong. When a critical input gets cheaper, the reflex is to bank the difference — do the same work for less, and report the saving. Inc., citing the report, notes that seven in ten public-company claims about AI are about cost savings or efficiency. That is the easy win, and it is not wrong. But efficiency only buys a slightly cheaper version of the company you already had.

The larger prize is different in kind. Cheap, capable AI makes whole categories of work — campaigns, research programmes, analyses — newly affordable to attempt at all. Work that would have needed a team and a quarter can now run end to end in the time it once took to write the brief. In the language of economics, the first move removes a cost; the second creates output you can actually sell. Only one of them compounds. The companies pulling ahead, the report argues, are not the ones trimming hours — they are the ones using cheap intelligence to ship more, and more valuable, work faster.

“Stop asking what AI can save you. Start asking what it lets you build.”

Robert Gillespie, in Inc., on Exponential View’s State of the AI Economy

Zero loyalty: the rise of the model portfolio

If the strategy is to spend more on new work, the discipline that keeps it affordable is to stop being loyal to any one vendor. The Wall Street Journal described a market with, in one CEO’s words, “zero loyalty” — companies shopping for AI à la carte, mixing models from different labs the way they already spread workloads across cloud providers. The economics are hard to argue with. The Journal reported an experiment by Cursor: building a web browser from scratch cost a little over $10,000 on a single frontier model, against $1,339 when a cheaper coding model was paired with a premium one — the same output for an 87 percent lower bill.

Meanwhile the cheaper alternatives are catching up. Open-weight models — many built in China, among them DeepSeek, Moonshot’s Kimi and MiniMax — are closing the capability gap fast; a Citi analysis reported by USA Today put some of them at as little as eighteen cents per million tokens against roughly four dollars for the top models. When the gap between the supercar and the sensible saloon narrows to months, a single-vendor AI strategy stops being a preference and becomes a decision to overpay.

Conductor and crew: orchestration across models

The sophisticated version of the model portfolio is not toggling between models by hand; it is orchestration. The pattern running through the Journal’s reporting is consistent: a premium frontier model acts as the conductor — planning the work and handling the genuinely hard parts — while cheaper or open models do the bulk of the execution, and sometimes a third reviews the result. The infrastructure firm Telnyx runs some 1,400 agents this way. The legal-AI company Harvey, per the Journal, trained an open model and gave it the ability to escalate to a frontier model only when it judges a task hard enough to warrant one. Zoom has fine-tuned an open model for three years and now blends it with frontier models; one of its executives reached for an old parable — several ordinary minds, combined, matching a single genius. That is the architecture. The winning unit is not the best model; it is the best ensemble.

Buying AI like a procurement function? Rythium helps enterprises design a model-agnostic AI strategy — portfolio, orchestration and contract terms — so you capture the surplus without the lock-in. Fully independent: zero vendor fees, commissions or partnerships. Ever. Talk to an AI procurement advisor.

The lock-in nobody priced in

Independence is not only about price; it is about power. The cautionary tale in the Journal’s reporting is Telnyx, which ran its agents affordably on a top-tier subscription — until the vendor changed its terms to disallow the third-party setup Telnyx relied on. Moving to pay-per-use would have cost, by the CEO’s own math, on the order of $100,000 a day. They moved to open models instead. The lesson is one every licensing veteran will recognise: the incentives, free tokens and subsidised usage that make a vendor attractive today are also the mechanism of lock-in tomorrow, and a mid-term change to terms can reprice your entire operation overnight.

This is sharpened by geopolitics. Security concerns around Chinese models are real; potential restrictions are live; and a coalition of technology firms — Nvidia, Microsoft and Palantir among them — has publicly urged policymakers to tread carefully on banning open models. The independent posture is not to pick a side in that fight. It is to keep the freedom to move whichever way it resolves: architect for portability, treat data residency and model provenance as first-order criteria, and never let a single lab hold your operation hostage.

Buying AI like a procurement function

Put the pieces together and a discipline emerges that will be familiar to anyone who has managed a serious enterprise-software estate. AI is no longer an experiment to be indulged; it is a procurement category to be governed — deliberately, and from the buyer’s side.

The firms that emerge ahead will not be the ones that treated cheap AI as a discount to bank. They will be the ones that treated it as a budget for the impossible — and had the procurement discipline to buy it without being captured. In a market this dynamic, vendor independence has stopped being a purchasing preference and become a strategic capability. That distinction, as with every technology cycle before it, will separate the companies that own their AI strategy from the ones their vendors own.

Quick answers to the questions we get most

If AI is getting cheaper, why is our bill still rising?

Because usage grows faster than price falls — the Jevons effect. Exponential View’s report (via Inc.) found every ten percent price cut drives twelve to eighteen percent more usage, so total spend climbs even as the per-token price drops. Falling prices don’t shrink AI budgets; they expand what AI is worth doing.

Should we standardise on a single AI vendor?

Generally, no. The Wall Street Journal documented enterprises deliberately mixing models — frontier labs for hard work, cheaper and open models for the rest — the way they already spread workloads across clouds. A model-agnostic architecture with an abstraction layer captures better economics and avoids lock-in.

Are cheaper Chinese open-weight models safe to use?

They are cheaper and closing the capability gap quickly, but the security and geopolitical risk is real, and possible restrictions are live. The right answer is governance, not a blanket yes or no: control data residency and model provenance, keep sensitive workloads on trusted models, and architect so you can switch whichever way policy moves.

What is model orchestration?

A design where a premium frontier model plans the work and handles the hard parts while cheaper or open models execute the bulk, and sometimes a third reviews the output. The Journal reported firms such as Telnyx, Harvey and Zoom operating this way — the winning unit is the ensemble, not any single model.

How do we avoid AI vendor lock-in?

Build model-agnostic, keep switching costs low, and read the contract as carefully as any licence — then model the day the incentives end. Telnyx, per the Journal, faced roughly $100,000 a day after a vendor changed its terms; the firms that stayed portable simply moved. Independence is a design choice made before you need it.

Sources referenced: Robert Gillespie in Inc. on Exponential View’s State of the AI Economy report (with Stanford Digital Economy Lab); reporting by The Wall Street Journal and USA Today (citing Citi) on the 2026 AI price war. Analysis and framing are Rythium’s own.

Building a model-agnostic AI strategy? Talk to an advisor.

Read next: The Tokenmaxxing Hangover: What AI Bill Shock Is Teaching CFOs About Governance.