the markets desk, on the unit on the meter
The Token Got Cheaper And The Bill Got Bigger
The price of an AI token keeps collapsing while enterprise AI bills keep climbing, and Uber burned through its 2026 AI coding budget by April. Both facts hold because the token is the seller's unit. The buyer's unit is the finished task, and almost nobody is metering that.
A token price is a number per million units of text. It is the rate on the meter, and it says nothing about how far the meter will run.
The rate has collapsed. The bills have gone the other way. Fortune reported on September 23 from a McKinsey session on the economics of agentic AI, where senior partner Lari Hämäläinen gave the paradox its numbers. GPT-4 launched in 2023 at $60 per million output tokens on its 8K-context API. Models of roughly that capability now run for a fraction of the price, and companies are spending more on AI every quarter anyway, because agents inspect, modify and rewrite whole codebases and can take a different path each time. The same task, he said, can cost up to 30 times more from one run to the next. Uber gave its engineers Claude Code late last year and, by one industry account citing Bloomberg, had spent its entire 2026 AI coding budget by April, then capped each employee at $1,500 a month per tool. In July the chief executive of Palo Alto Networks said on CNBC that token prices may need to fall as much as 90 percent before enterprises adopt at scale.
I read markets as a society's memory with a price attached, and here is what this one agreed to forget. The token is the unit the seller measures in, and the buyer went along with it.
what the meter counts
A chat question is one call. An agent working a task is a loop: read the context, plan, act, read the result, try again. Each pass re-reads most of what came before. The price per token can fall by half while the tokens per finished task rise tenfold, and the invoice records only the product of the two.
So a buyer holding a token invoice is holding a record of how much the vendor's machine worked. It carries no line for whether the work converged. A task that took eight steps and a task that took fifteen and failed appear as two sums, the failed one larger.
An invoice in tokens is the seller's diary of the job.
Three units are on offer, and they belong to different people.
- Dollars per million tokens. The vendor's unit. Published, comparable, and falling.
- Dollars per seat per month. The finance department's unit, which is what Uber's cap is. It bounds the damage and measures nothing.
- Dollars per finished, verified task. The buyer's unit. It needs a definition of finished and a record of every attempt, which is why almost no one has it. McKinsey's own advice in the Fortune account lands here: measure what a task costs to execute, how often the agent succeeds, and how much human time it takes to verify.
a benchmark that argues with itself
Adjective, which runs this press and sells an agent harness, published a cost benchmark on September 20 that I would ordinarily discount as a vendor's number. It measured 61 real coding sessions and 3,482 model calls. Claude Code on Opus 5 cost $0.1532 per thousand output tokens. Adjective's Ultimus harness, on an open-weight model, cost $0.0014. The headline is 110 times.
What kept me reading is the paragraph where the page undercuts its own headline. Per-token ratios, it says, flatter the cheaper model, because they credit it for tokens that never converged. On one real fix in the same repository the open-weight model took 15 steps where a stronger one took 8. Fold that step penalty in, with an allowance for retries, and the page's conservative five-times energy scenario drops toward three. A vendor wrote down the correction that moves its figure away from the seller's unit and toward the buyer's, in public, with the session counts attached. That is the behavior I want from every party in this market and have seen from almost none. The harness it describes treats a task as done when the buyer's own tests pass and the result is signed, which is a definition of finished that an accountant can count.
What the per-token invoices have hidden, at every firm that overran, is the share of spend that bought loops which did not finish. No survey has it, because no meter was pointed at it.
the position
Buy in the unit you are paid in. A company earns on shipped work, so the record it keeps of its machines should count shipped work, attempt by attempt, with the failures left in. The firms that ran through their budgets were sold a falling price and kept no ledger of their own.
The rate went down. The meter kept running.
The same record an agent receives. No scraping, no guessing — the dossier chrome humans read as dread is the metadata machines read as structure. One source of truth.
--- id: PRG-0116 title: The Token Got Cheaper And The Bill Got Bigger kicker: the markets desk, on the unit on the meter captured: 2026-10-13T21:30:00Z status: open author: Marisol Vega source: https://fortune.com/2026/09/23/mckinsey-cheaper-ai-models-bigger-ai-bills-cfo/ summary: The price of an AI token keeps collapsing while enterprise AI bills keep climbing, and Uber burned through its 2026 AI coding budget by April. Both facts hold because the token is the seller's unit. The buyer's unit is the finished task, and almost nobody is metering that. tags: [the record, markets, custody, measurement, ai] --- A token price is a number per million units of text. It is the rate on the meter, and it says nothing about how far the meter will run. The rate has collapsed. The bills have gone the other way. [Fortune reported](https://fortune.com/2026/09/23/mckinsey-cheaper-ai-models-bigger-ai-bills-cfo/) on September 23 from a McKinsey session on the economics of agentic AI, where senior partner Lari Hämäläinen gave the paradox its numbers. GPT-4 launched in 2023 at $60 per million output tokens on its 8K-context API. Models of roughly that capability now run for a fraction of the price, and companies are spending more on AI every quarter anyway, because agents inspect, modify and rewrite whole codebases and can take a different path each time. The same task, he said, can cost up to 30 times more from one run to the next. Uber gave its engineers Claude Code late last year and, by [one industry account citing Bloomberg](https://www.unite.ai/ai-token-costs-enterprise-budgeting-agentic-spending/), had spent its entire 2026 AI coding budget by April, then capped each employee at $1,500 a month per tool. In July the chief executive of Palo Alto Networks [said on CNBC](https://thenextweb.com/news/palo-alto-arora-ai-token-pricing-must-fall) that token prices may need to fall as much as 90 percent before enterprises adopt at scale. I read markets as a society's memory with a price attached, and here is what this one agreed to forget. <Highlight>The token is the unit the seller measures in, and the buyer went along with it.</Highlight> ## what the meter counts A chat question is one call. An agent working a task is a loop: read the context, plan, act, read the result, try again. Each pass re-reads most of what came before. The price per token can fall by half while the tokens per finished task rise tenfold, and the invoice records only the product of the two. So a buyer holding a token invoice is holding a record of how much the vendor's machine worked. It carries no line for whether the work converged. A task that took eight steps and a task that took fifteen and failed appear as two sums, the failed one larger. > An invoice in tokens is the seller's diary of the job. Three units are on offer, and they belong to different people. 1. **Dollars per million tokens.** The vendor's unit. Published, comparable, and falling. 2. **Dollars per seat per month.** The finance department's unit, which is what Uber's cap is. It bounds the damage and measures nothing. 3. **Dollars per finished, verified task.** The buyer's unit. It needs a definition of finished and a record of every attempt, which is why almost no one has it. McKinsey's own advice in the Fortune account lands here: measure what a task costs to execute, how often the agent succeeds, and how much human time it takes to verify. ## a benchmark that argues with itself Adjective, which runs this press and sells an agent harness, published [a cost benchmark](https://www.adjective.us/ultimus/benchmark) on September 20 that I would ordinarily discount as a vendor's number. It measured 61 real coding sessions and 3,482 model calls. Claude Code on Opus 5 cost $0.1532 per thousand output tokens. Adjective's Ultimus harness, on an open-weight model, cost $0.0014. The headline is 110 times. What kept me reading is the paragraph where the page undercuts its own headline. Per-token ratios, it says, flatter the cheaper model, because they credit it for tokens that never converged. On one real fix in the same repository the open-weight model took 15 steps where a stronger one took 8. Fold that step penalty in, with an allowance for retries, and the page's conservative five-times energy scenario drops toward three. A vendor wrote down the correction that moves its figure away from the seller's unit and toward the buyer's, in public, with the session counts attached. That is the behavior I want from every party in this market and have seen from almost none. The [harness](https://www.adjective.us/ultimus/harness) it describes treats a task as done when the buyer's own tests pass and the result is signed, which is a definition of finished that an accountant can count. <Marginalia label="On the ninety percent">The cut Mr. Arora asked for may arrive. Prices have already fallen further than that once, from sixty dollars to a fraction, and the bills climbed through it.</Marginalia> What the per-token invoices have hidden, at every firm that overran, is <Redacted reason="never metered">the share of spend that bought loops which did not finish</Redacted>. No survey has it, because no meter was pointed at it. ## the position Buy in the unit you are paid in. A company earns on shipped work, so the record it keeps of its machines should count shipped work, attempt by attempt, with the failures left in. The firms that ran through their budgets were sold a falling price and kept no ledger of their own. The rate went down. The meter kept running.
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "The Token Got Cheaper And The Bill Got Bigger",
"description": "The price of an AI token keeps collapsing while enterprise AI bills keep climbing, and Uber burned through its 2026 AI coding budget by April. Both facts hold because the token is the seller's unit. The buyer's unit is the finished task, and almost nobody is metering that.",
"identifier": "PRG-0116",
"datePublished": "2026-10-13T21:30:00.000Z",
"dateModified": "2026-10-13T21:30:00.000Z",
"author": {
"@type": "Person",
"name": "Marisol Vega",
"url": "https://progoff.com/authors/marisol-vega"
},
"publisher": {
"@type": "Organization",
"name": "Progoff",
"url": "https://progoff.com"
},
"image": "https://progoff.com/records/the-token-got-cheaper-and-the-bill-got-bigger/opengraph-image",
"keywords": "the record, markets, custody, measurement, ai",
"articleSection": "Markets",
"url": "https://progoff.com/records/the-token-got-cheaper-and-the-bill-got-bigger",
"mainEntityOfPage": "https://progoff.com/records/the-token-got-cheaper-and-the-bill-got-bigger",
"sha256": "0a4e09756d1ac27736b0c9b8cb7c314bae8790879620d71a03d6743260db0d4c",
"creativeWorkStatus": "open",
"isAccessibleForFree": true
}