The strongest reply is that this compares a thin 2023 reply with a thick 2026 agentic job and calls the difference inflation. Hold the task and the quality bar fixed, and the price of that frozen answer falls with the token rate. That is essentially Payden & Rygel’s case: adjusted for quality, AI is deflationary.
The token price really is falling
This is not only a forecast. Uber exhausted its full-year AI budget in the first four months of 2026, its chief technology officer said in April. In June, it capped employee token spending at ,500 a month per agentic coding tool, Bloomberg reported. J.P. Morgan’s July analysis, citing Ramp data, found the top 1 percent of AI users consuming more than 600 times as many tokens as the median user, with that intensity skewed toward complex coding work. Those token-heavy workflows are exactly where the cost of an answer climbs fastest.
Gartner expects the gap to widen. In an August 17, 2026 forecast, it predicted that inference costs per agentic workflow will rise more than fivefold through 2028. It also estimated that routing a task to an agentic reasoning model already raises provider inference costs at least five times over a basic chatbot interaction. Gartner calls this the Inference Paradox: better unit economics pushing total AI costs up without a reliable path to matching value. “Product leaders cannot rely on more efficient token economics to rationalize AI costs,” said Will Sommer, a senior director analyst at Gartner.
An answer is no longer one reply
The objection is right about frozen tasks. A one-shot prompt on a model you never upgrade is a genuine bargain. But companies are not buying frozen tasks. They are moving work onto agents precisely because agents do more, and Gartner’s forecast is measured per agentic workflow, the unit companies now deploy. A better answer can be worth the money. It is still a more expensive answer, and a falling rate card makes that easy to miss.
The price of an AI token has collapsed. The cost of getting finished work out of AI has not. The two are moving in opposite directions, and the first is hiding the second. A token is a rate card. A completed task, once an AI system plans, calls tools, checks its work and tries again, is a loop that can burn dozens of times more tokens than a single reply. Finance teams that celebrate the falling rate are watching the wrong number.
The bill is already arriving
The unit that delivers work in 2026 is not a single completion. Data from Wolfe Research, published by J.P. Morgan in July 2026, puts a single chatbot prompt at roughly 900 tokens and analyzing and debugging code at roughly 56,000, more than 60 times as many. The FinOps Foundation, the industry body for cloud financial management, estimated in May 2026 that reasoning and agentic workloads use five to 30 times more tokens per task than the equivalent chat interaction. It also found that per-token price declines have slowed and now sit mostly in commodity tiers rather than at the frontier. Its summary: “A token may get cheaper; tokens, in aggregate, are not.”
The fair objection
A procurement review that stops there will report a win.
For a fixed level of capability, the decline is not in dispute. Asset manager Payden & Rygel calculated in May 2026 that the cost of a million tokens from models matching GPT-4’s level of capability fell more than 99.9 percent, from at launch to Model APIs are sold per million tokens, which is why price indexes and finance reviews track that unit. The FinOps Foundation’s FOCUS specification normalizes cost and usage data across providers so token spend can be compared, but it does not define a cost per completed task. In June 2026, the Linux Foundation announced its intent to form a Tokenomics Foundation to set open standards for AI cost management. That effort is the natural place to fix the blind spot. Routing, caching and smaller models can cut the tokens used per call, but no public figure yet shows those savings outpacing the growth in tokens per task.
What would prove this wrong
This argument fails if the cost per successfully completed agentic task falls in step with per-token prices through 2028, instead of rising as Gartner forecasts. It also fails if measured tokens per task stop growing as agents mature. Either result would show that cheaper tokens are, in fact, producing cheaper answers.
What finance should publish
Finance leaders who buy model access should report, every quarter, cost per successfully completed task next to cost per million tokens, using one task definition they do not quietly change. The check is simple: if the per-task figure is missing from the review, the review has not measured what AI actually costs. The Tokenomics Foundation should make cost per completed task a core unit in its first standard. Whether that unit appears in the published specification is the test.
