This is part of WeGoAI's daily series on the things that actually shape our world. Figures from the Bargo Token Demand Index, American Compute's Open-Weight LLM Price Index and InferIndex as of 19 October 2026. Not investment advice.

“This is part of WeGoAI's daily series on the things that actually shape our world.”

Volume and Value Have Come Apart

The AI industry spent three years arguing about whether inference demand would show up. On 19 October the answer was in the tape: 136.1 trillion tokens per week of tracked inference, a demand index reading of 578.9, up 31% in thirty days and 479% since 6 May 2026.

The price attached to that volume is the interesting part. The demand-weighted effective price is $0.82 per million tokens. Open-weight models carry 82.1% of tracked volume at a blended $0.42 per million; the closed providers — Anthropic, OpenAI, Google and xAI together — carry 12.8% of the volume at $3.35 per million.

Put those together and the market splits cleanly in two. Cheap open weights are the volume business; closed frontier models are the value business. On list prices, closed models account for roughly 52% of the spend on 13% of the tokens — and they are now the only part of the market that is getting more expensive.

A caveat that matters more than the headline: this index is built from OpenRouter's trailing-seven-day rankings, a large but partial slice of the market. Direct enterprise contracts, cloud marketplace deals and self-hosted deployments are outside it. Read it as a weather station, not a census. It is also genuinely volatile — the prior day's reading was 142.9T tokens at an index of 607.6, so the weekly figure wobbled 4.8% in a single day. Trends here are real; single points are not.

What the composition does show is how concentrated cheap demand has become. DeepSeek V4.1 Flash alone accounts for 29.6T tokens a week — 26.5% of the entire open-weight group, all by itself. GLM 5.3 Flash adds 11.9T, MiMo-V2.6-Flash 9.6T, Step-5-Preview 8.2T, Hunyuan Hy4-Preview 7.8T. Five Chinese labs' models are the load-bearing wall of the world's cheap inference.

The Frontier Rotates Faster Than It Grows

The other side of the market is where the argument is. In the last thirty days, the Anthropic group's tracked volume grew 225% while the OpenAI group's fell 43%. Those are not small numbers, and they are moving in opposite directions inside a market that grew 31% overall.

Two readings are possible, and both are probably partly true. The first is rotation: model-router traffic is the most price- and benchmark-elastic demand in the industry, and when a new frontier release wins an eval cycle, the traffic moves within days. The second is mix: the OpenAI group's blended price is $3.25 per million against Anthropic's $5.00, which is consistent with OpenAI holding more of the high-volume, low-margin commodity work while Anthropic took the agentic, long-output, high-price work that actually looks like margin.

Either way, the practical lesson for anyone buying inference is uncomfortable: on a router-visible market, being the best model is worth a three-digit percentage swing in volume in a single month, in both directions. Distribution and switching costs are doing much less work than the labs' marketing implies.

Nine Prices for One Model

Now the part that has changed most in 2026, and that almost nobody has priced into their plans.

Take DeepSeek V4.1 Flash — again, the single most-used model in the index. Twenty-nine providers sell it. Output tokens list between $0.18 and $1.60 per million, around a median of $1.00. That is an 8.9x spread on identical weights. Nothing about the model differs between those endpoints; only the host, the region, the rate limits and the sales motion.

The dispersion is not a DeepSeek quirk. Llama 3.1 8B lists at a 7.2x spread across five providers, GLM 5.3 Flash 6.4x, Mistral Nemo 6.3x, gpt-oss-120b 5.6x. The tightest families — Kimi K3 at 1.4x, MiMo-V2.6-Flash at 1.6x — are the ones with the fewest sellers.

And the gap is widening rather than closing. Over the 30 days to 9 October, the cheapest output prices in the index fell 35%, while the median fell 6%. Competition is happening at the bottom of the book, where aggressive new capacity is dumping. The middle of the book is sticky, because that is where buyers stop shopping.

Look at the bottom and it looks like a rout: Mistral Nemo's median output price is $0.03 per million, down 68% in thirty days. Nemotron 3.5 Lightning fell 35% to $0.147. Look at the top and it barely moves: Kimi K3 still lists at a $14.00 median across 20 providers, with a $9.60 floor. Cheap and expensive models are not converging; the same model is available at both ends of the book.

The Index Assumes You Shop

The cleanest weekly cross-section comes from InferIndex, which publishes every Monday at 00:00 UTC and defines a family's level as the median across models of the median of each model's three cheapest providers. In other words, the index is already measuring the effect of shopping around.

On 19 October the market levels per million tokens (3:1 input-to-output blend) were:

  • Frontier closed: $3.75 — up 3.4% week on week, and the only family that rose.
  • Code: $0.744 — down 7%.
  • Compact closed: $0.513 — down 8.9%.
  • Open weights, 200B+: $0.45 — unchanged.
  • Open weights, 40–200B: $0.329 — up 11.4%.
  • Open weights, under 40B: $0.128 — up 2%.

Read that table as a sequence of decisions rather than a scoreboard. Frontier closed pricing drifts upward, because the labs selling it compete on capability and can. The open middle — the 40B-to-200B band that a competent team can actually serve on its own hardware — is where price moves most, in both directions, because capacity keeps arriving. And the small-model band sits at 12.8 cents per million: that is not a price, that is a rounding error, which is why the interesting architectural question in 2026 is not "which model" but "which model for which step of the pipeline".

What This Means

For buyers. The marginal cost of a token is now set by procurement, not by a lab's price list. Choosing the median provider instead of the cheapest costs you 4x to 9x on identical weights, forever, invisibly, because nobody benchmarks a spreadsheet against a router. Model routing and provider arbitrage are the highest-return infrastructure work available right now — and they are cheap to build compared with what they save.

For labs and providers. Volume share and revenue share have decoupled. Serving 82% of the world's tokens at $0.42 per million is a real business — throughput, utilisation, depreciation — but it is a very different business from selling $5.00-per-million agentic work. Any provider that cannot say which of those two it is in will be squeezed by someone who can.

For the infrastructure story. Yesterday's post argued that only about a third of the announced data-centre pipeline is credible. The token data is the other half of that argument, and it cuts both ways. Cheaper inference is exactly what makes a gigawatt of committed capacity look underwriteable — demand is compounding at 479% since May and the marginal token funds itself. But it also means the revenue per megawatt falls every quarter. The buildout is a bet on volume, not on price, and the two-sided split in this chart is what that bet looks like from the inside.

The risk case. A market where one group supplies 82% of volume at 42 cents, where thirty-day swings in either direction are normal, and where the cheapest prices fall 35% in a month is a market with no floor under unit economics. If open-weight capacity keeps arriving faster than demand compounds, the effective price goes to the cost of electricity plus depreciation — and the frontier's pricing power narrows to whatever genuinely cannot be replicated by open weights. That is the number to watch, not the headline token counts.

Bottom Line

Demand is compounding, prices are not. One model now sells at nine prices depending on where you buy it, the same open weights sit at both the floor and the ceiling of the book, and the cheapest 35% of the decline in output pricing accrued only to buyers who went looking for it. The inference market is no longer a market for models. It is a market for procurement, and it is rewarding the people who treat it that way.

---

Sources: Bargo Token Demand Index (OpenRouter trailing-7-day token counts and blended median list prices, snapshot of 19 October 2026); American Compute Open-Weight LLM Price Index (daily list prices through a large public router, 12 July – 9 October 2026); InferIndex weekly market index, published 19 October 2026. Routers are one channel among many; direct and enterprise pricing is excluded. Not investment advice.