Model buying is two markets:
volume vs money

The cheapest model is not the winner — the model that fits the task is. Vercel’s September AI Gateway Production Index is the first neutral, public time-series of what production AI actually costs, and it shows a structural split: open-weight models ran 56% of gateway tokens in August but took only ~14% of spend; nine in ten teams that used Anthropic’s top model stepped down to the tier below; and the average token price halved in five months. Our judgment: “average token price” is the wrong unit of purchase. The decision unit is model tier per task — and the numbers to trust are the ones you measure on your own workloads.

1. Volume and money are splitting into two markets

In August 2026, open-weight models ran a majority of tokens on Vercel AI Gateway for the first time: 56%, up from 7% in December 2025 — eight months from under one in ten to more than all closed-weight models combined. But volume is not money. The frontier kept the majority of spend, and open-weight dollar share, while accelerating, is still a minority. Anthropic alone has taken at least 61 cents of every dollar spent through the gateway every month since December, and 64 cents in August. Source: Vercel’s report.

What this shows is not “open source is cheap” — it is that two markets are forming: a commodity volume layer (open weights) and a quality-premium money layer (frontier). For C-suite the implication is direct: the “we went open weight so we saved money” narrative is wrong — the money is still in the frontier. The real question is whether each task tier’s spend is justified by its output.

2. Nine in ten teams stepped Fable 5 down to Opus 5

The cleanest proof comes from inside one lab. When the US export control on Fable 5 was lifted on July 1 and access restored, its gateway spend share surged to 13.2%. At the end of that same month Opus 5 came online. In August, Fable 5’s spend share fell to 4.9% while Opus 5’s rose to 22.5% — nine in ten teams that ran Fable cut their usage, and more moved to Opus 5 than to any other model. Vercel’s conclusion in one line: Fable’s extra capability wasn’t worth double the price.

The detail worth noting: the workload stepped down, but the money did not leave Anthropic. Teams went to Opus 5, not to another lab. That is the second pattern — customer loyalty follows the model profile, not the brand: a new model that preserves what users valued in its predecessor keeps them; one that doesn’t loses them to other providers.

3. Average token price: read the trend, not the number

The average token price fell 23.2% month over month in August — the third straight drop and the steepest since April — and costs less than half what it did five months ago. Among teams running more than ten million tokens in both months, the median team paid 7.6% less. For budgets, this is the good trend: the same budget buys more inference, and frontier models can be reserved for the tiers that justify the premium.

But the single number misleads. First, the methodology: spending is estimated at labs’ published list prices, actual bills may differ, and the report notes methodology is revised. Second, the average hides the structure: Google’s token share fell from 30% to 5%, with 22 of the 25 percentage points coming from Gemini 3 Flash alone — half its volume went to cheaper models like GPT-5.6 Luna, most of the rest to Opus 5 and Sonnet 5 at roughly nine and three times the price. The average barely moved; the decisions changed completely.

4. Our judgment: the purchasing unit is model tier per task

“Average token price” invites the wrong decision: pick the cheapest. But the 9-in-10 Fable-to-Opus step-down shows real buyers already optimize per task — they want “good enough”, not “most expensive” and not “cheapest”. Cost control was never about choosing the cheapest model; it is about choosing the right tier — reserve the frontier budget for loads that can prove the premium, put commodity volume on open weights, and back it with a quality gate.

Our position: delete “average token price” from your budget file and replace it with “cost per completed task by tier”. This is not a wording dispute — the 56%/14% split means “we went open weight so we saved” and “we picked the right tier for each task” are different claims, and only the second is real cost control. It is also how we run our own stack: route by task tier, measure by completed task, not by token.

5. The measurement caveat is the point

Vercel is the closest thing to a neutral observation point in public data, but it is still a gateway operator, and the report itself states: spending is a list-price estimate, actual bills differ, methodology is revised, and the comparisons track relative movement among the same teams rather than individual tokens. These caveats are not flaws — they are the argument: vendor-adjacent telemetry is not an audit.

For AI app leads: instrument per-task cost and quality in every tier; that is the routing signal. For CFOs: read “average token price” as a trend indicator, not a budget line. The bill-level numbers you can trust are the ones you measure on your own traffic.

6. Buyer checklist and the question left for you

Action: reserve frontier spend for loads that can prove the premium; put commodity volume on open weights behind a quality gate; and start building task-level cost telemetry in your stack today. The question left for buyers: if you replaced every “average token price” in your budget deck with “cost per completed task by tier”, would any decision change? If not, you are not buying by fit yet.

References: Vercel, AI Gateway Production Index — September 2026 (spending estimated at list prices; methodology caveat stated) https://vercel.com/blog/ai-gateway-production-index-september-2026 ｜ The New Stack, open-weight volume vs Anthropic spend https://thenewstack.io/open-weight-anthropic-spend/