The US AI Labs, Part 3: One Token, Two Markets

On 26 October Silicon Data's token index, the series Bloomberg carries as SDLLMTK, printed 92 cents per million tokens. It weights each listed price by the volume actually used, and it is less than half the $2.07 it reached in late May.

That same week Claude Opus 5.5's list prices blended to $8 a million, and Anthropic's registration statement reported $13.1 billion of first-half revenue metered in tokens, 80% of its total.

The three figures describe a market that has come apart. The average is falling because buyers moved. The top of the range is not falling at all. And the companies with the largest revenue are selling the top.

This part prices what they sell in a unit sized for the business. A boken is one billion tokens; the name is ours, not an industry standard. Per-million prices suit one developer's month, but the industry's volumes run to quadrillions, where per-million figures stop comparing. Per boken they read like barrels.

A boken of the frontier costs thirty times a boken of the commodity

At list, a boken of Claude Opus 5.5 costs $8,000: $4 per million on the way in, $20 out, blended three to one. A boken of DeepSeek V4.1 Flash costs $262.50, on 15 cents in and 60 cents out. Opus scores 58 on the Intelligence Index and Flash 32, so Opus is clearly the stronger model, but the price of their volume differs 30 times over.

The ladder in the table shows where that difference sits. It is not a smooth gradient from cheap to capable.

Read down the price column at a fixed score. DeepSeek V4.1 Flash and GLM-5.3 both score 32 and charge $262.50 and $1,259 a boken. Gemini 4 Argon and Claude Fable 5.1 both score 50, at $4,000 and $20,000. Claude Sonnet 5.5 at 54 and GPT-6.1 Sol at 55 both list at $4,000. Capability explains some of that ordering and almost none of its magnitude.

GPT-6.1 Sol is the clearest case. It scores 55, three points below Opus, and a boken costs $4,000 against Opus's $8,000. On the same evaluation it costs $1.27 per task against Opus's $10.04, about eight times less. That is a far wider gap than the boken prices imply, because Opus also spends more tokens getting to an answer.

A boken measures volume. A customer buys a finished task. The two diverge by however long a model thinks before it answers, which is why a price list and a bill can tell different stories about the same model.

The unit has a second limit. Models split text into tokens differently, so a billion tokens of one model is not a billion tokens of another in pages read or code written. Every figure in this piece therefore names its model. A boken compares prices; it does not compare work.

The average halved because volume moved, not because the price list did

OpenRouter, a public router whose rankings show which models developers actually pay for, carried 190 trillion tokens in the week to 26 October, by Bargo's count, up 708% since early May. Open-weight models carried 74% of them. The single busiest model was DeepSeek's V4.1 Flash 1017 at 35.4 trillion tokens; Claude Opus 5.5, one of the most expensive models on the ladder above, carried 8.2 trillion.

Part 2 ended on the risk that a token price falling faster than usage grows would leave the same capacity earning less. This year it went the other way: Silicon Data's price halved from its May peak while routed volume grew eightfold, so spending at those prices rose about 3.6 times.

Those tokens are cheap. Open weights blend at $0.42 per million against $3.67 for the closed labs, so the routed market pays for volume in one place and for capability in another. Open weights take 74% of the tokens and 28% of the money. The four closed labs take 21% of the tokens and 68% of the money.

An average over those two groups is a statement about the mix, not about anyone's price.

That is what the index records. Silicon Data's series is down by more than half from its May peak, while BenchLM's frontier basket, which tracks flagship models rather than volume, sits at a median $5.63 per million and 61% above a year earlier. Over three years the frontier is still 85% below its March 2023 level. Deflation is real; it is the last twelve months that run the other way.

Both indices are honest. They disagree because one asks what a token costs and the other asks what buyers paid, and in 2026 the answer depends on which buyers you count.

The premium half is profitable, and it is priced off a handful of index points

Anthropic sells that half. In the first six months of 2026, $13.1 billion of its revenue, 80%, was metered in tokens, and $3.2 billion came from subscriptions. At the $5 per million its traffic fetches on OpenRouter, the metered figure implies about 2.6 quadrillion tokens in six months.

Anthropic's compute bill grew much more slowly than its revenue. Compute and infrastructure went from 201% of revenue in the first half of 2025 to 55% in the same half of 2026, and to 48% in the June quarter, which produced $974 million of operating income, the first operating profit the filing reports. Net revenue retention was about 900% at the end of September.

On those numbers the premium is not a mark on a chart. It is a margin, earned by selling tokens at eight to twelve times the commodity price while compute takes a shrinking share of every dollar.

What holds the margin up is capability. The closed labs' 68% of spend is paid for models that lead the index. Their 21% share of volume is the price of that lead: buyers route around it wherever a cheaper model is good enough. The pressure on that lead comes from inside the closed group. GPT-6.1 Sol scores 55 for half Opus's boken price; the best open-weight models on the current index, GLM-5.3 and DeepSeek V4.1 Flash, score 32. If the next release closes the remaining gap the same way, the premium stops being a rent on capability and becomes a rent on the last few points of it.

The case against that reading is that the premium has held so far. Closed models still take two thirds of the money on a fifth of the tokens, frontier prices rose this year, and 900% retention, which excludes Pro and Max subscribers, says Anthropic's API and enterprise customers are not shopping on price. All of that is true, and all of it describes a rent: rents are paid until the substitute arrives.

Two disclosures would settle it. A model scoring 57 or higher below $2,000 a boken would mean the premium tracks production cost, and production costs are still falling; today the best score at that price is GPT-6.1 Terra's 51. Closed labs holding two thirds of spend a year from now, while open weights keep taking volume, would mean the split is stable and the rent is structural.

Anthropic's filing sets 900% retention beside a token price that halved. Both can hold for a while. They stop holding together if tokens get cheaper faster than the tasks get harder.

Three parts, one bet on a few index points

Read together, the series describes one wager. Part 1 found Anthropic valued at 21 times the run-rate it claimed in May and 210 times its 2025 revenue, a price that only makes sense if the revenue line keeps climbing. Part 2 found the suppliers paying for much of that climb and taking their money back through contracts: Anthropic's three cloud agreements alone commit it to $269.4 billion, about $33 billion a year, priced in capacity rather than tokens.

This part found what the climb is made of. Of Anthropic's first-half revenue, 80% was metered in tokens, sold at the top of a market whose average price halved this year. On OpenRouter, Claude traffic fetches $5 a million and open weights 42 cents. Anthropic sells at the expensive end of that gap, and in the June quarter compute took 48% of its revenue and it made an operating profit.

The valuation, the supplier money and the take-or-pay contracts all assume the frontier keeps its price.

That price rests on a narrow lead. GPT-6.1 Sol scores three points below Opus for half its boken price and about eight times less per task. The nearest challenger is a rival closed lab, not an open-weight model.

What to take from the series:

• The valuations price a slope, and the slope is metered tokens. With 80% of revenue scaling with token price and volume, a 21x run-rate multiple is a bet on what a frontier token keeps selling for.
• The lab carries the price risk, not its suppliers. Anthropic owes about $33 billion a year in cloud commitments whether tokens get cheaper or not. A true-up payment or a drawdown on the Broadcom notes would be the first sign the risk is landing.
• For buyers, the discount sits just below the top. The model three points under Opus costs half as much per boken. Whether the last points are worth twice the price is now a choice, not a given; the effort-level brief makes the same case one setting down.
• Watch the filings, not the next funding round. If the listing goes ahead, each quarterly report will show whether compute keeps falling as a share of revenue while token prices fall. That one line answers the question all three parts ask.

None of this says the bet fails. Anthropic's first half shows the structure can pay: revenue 14 times the year before, compute down to 48% of it by June, a first operating profit. It says the bet is narrower than a $965 billion valuation makes it look. Everything in this series rests on a lead you can measure in index points, and any model release can move it.