The token/revenue split: open weights took 78.4% of the volume and 5% of the money
Date: 2026-09-20
Topic: ai-industry
Source: Vercel AI Gateway data, reported on the X home feed (@MelvinInvests, @rohanpaul_ai)
Raw: raw/twitter/feed/2026-09-20-morning-ranked.json
TL;DR
Usage data from Vercel's AI Gateway, a router that sits in front of many model providers and therefore sees a real cross-vendor traffic mix, shows a split that has gone from notable to structural in three months.
By token volume: open-weight models 78.4%, closed models 21.6%. In June the open share was roughly 40%, so it has nearly doubled in one quarter. DeepSeek V4.1 Flash alone is 59.3% of all tokens crossing the gateway. Claude Opus 4.8 and Claude Sonnet 5 are 1.7% each.
By spend, the picture inverts. Claude Opus 4.8 takes 13.7% of dollars on 1.7% of tokens. Anthropic's five listed models together take 36.9% of total spending. DeepSeek V4.1 Flash moves most of the tokens and collects 5.1% of the spending.
So the two businesses are no longer the same business. Open weights own the volume tier; closed models own the revenue tier by charging roughly an order of magnitude more per token for the work people will not delegate downward.
Why this is a routing result, not just a market result
This is the clearest production-scale confirmation to date of the premise every paper on the routing page assumes and none of them measures: that real traffic is sharply separable into a large cheap tier and a small expensive tier, and that operators will act on the separation once the cheap tier is good enough. A gateway is a router. The 78.4/21.6 volume split against the 5.1/36.9 spend split is an empirical picture of what a large population of developers decided the value-of-information trade actually is, without anyone writing the objective down.
It also sharpens a distinction the wiki has been sloppy about. Most routing papers optimise accuracy per dollar on a benchmark. This data says the deployed behaviour is closer to a floor than an optimum: send everything to the cheap model unless the task is one you have already decided needs the expensive one. That is a much coarser policy than any router in the literature, and it is capturing most of the available saving. The marginal value of a sophisticated router is whatever is left after the coarse policy runs, and nobody has measured that residual.
One vendor at 59.3% of volume is its own risk finding. A gateway whose traffic is majority one open model is not a diversified open ecosystem, it is a single point of dependence with a permissive licence. The usual argument for open weights is that you can self-host, fine-tune and move clouds. That argument holds for the licence and not for the concentration.
What it does to the industry thread
It gives Lambert's compute-allocation worry a mechanism. The RSI entry (09-20) records Nathan Lambert arguing that frontier labs may not be able to hold a constant share of compute on internal R&D as volume grows, especially under IPO scrutiny. If the volume tier is migrating to open weights while the revenue tier holds, then the labs' revenue is increasingly concentrated in a narrow band of hard tasks. That is a higher-margin business and a smaller one, and it is more exposed to the open models improving by one notch than a volume business would be.
It reframes the week's regulation fight. Today also brought a class action in California federal court alleging that the September 12 call for industry-wide coordination to limit AI progress, plus same-day supportive statements from three rival labs, constitutes a Section 1 Sherman Act restraint on output. The most-shared framing on the feed connects the two directly: the argument runs that a slowdown pact conveniently freezes the capability gap at the moment open models are closing it from below. That is a motive claim and not evidence, and it should be labelled as such. The gateway data is evidence about market structure; the inference about intent is commentary.
It is consistent with the decision-model story. The other dominant thread today, the Jev ecosystem census (09-20), is 160 projects built in 72 hours around pushing the cheapest class of decisions off the frontier model entirely. Same direction, one tier further down: first the bulk generation moves to open weights, then the bulk decisions move to a model that does not generate at all.
Gaps
One gateway is one population. Vercel's users skew toward web and agent developers building cost-sensitive products, which is exactly the population most likely to route downward, so the 78.4% is an upper bound on the general market rather than an estimate of it. Token volume is also a poor proxy for work done: reasoning models emit many internal thinking tokens per unit of output, so a share measured in tokens flatters whichever models think longest. No per-task or per-request breakdown is published. And the June comparison point is quoted without a source, so the doubling claim rests on the same single report.