The ratio behind the 29% share is the interesting number: nearly a third of tokens for just under 4% of spend [1] works out to roughly an eighth of the average dollar-per-token on the gateway, and Vercel puts open-weight pricing at about a tenth of the platform average [8]. That is a discount deep enough that a buyer moving a workload does not need a procurement argument, only a routing rule.
Which is why the flat price per token [2] is the line worth pausing on. Cheap volume tripling its share since April [7] should have dragged the blended average down, and it would have, except closed-weight frontier prices rose about 12% per token in the same month [9] and cancelled it out. The average held still because two things moved in opposite directions, not because anything settled. Strip out the frontier increase and the same June looks like deflation; strip out the open-weight growth and it looks like a price rise. Same month, three readings.
Anthropic's position is the part I would not short. It took 61% of spend on 32% of tokens [3], down from 65% in May but in line with April [11], and 72% or more of spend in coding agents, back-office agents and app generation [4] - the work where a wrong answer costs real money. If a buyer's cheap tier is absorbing summarisation and classification, the frontier bill is not being cut, it is being protected. The pattern with OpenAI is the cleaner tell: token share fell from 12.5% to 10.3% while spend share rose from 13.3% to 16.1% [12], which Vercel reads as cost per token up about 50% relative to the market in a month [13]. Less volume, costlier work.
This is probably wrong, but I read the DeepSeek climb to 22.6% of tokens, within two points of Google [5], as less about model quality than about the absence of any switching cost worth naming, and GLM 5.2 is the evidence: MIT-licensed, priced near a fifth of Opus 4.8, 50x daily token growth from June 16 to month end, and 76% of its own family's June tokens in barely two weeks [14][15], against Gemini 3.1 Pro needing a second month to do the same [16]. Or rather, the more interesting version of the claim is that adoption is now faster than evaluation, since Vercel notes customer count grew far faster than token usage [17] - people are trying it before they trust it with anything.
This same month reads differently depending on which thread you pull. The frontier price increases could be temporary, in which case the blended price falls hard once they stop and the volume story becomes a straightforward cost win. Or only about one in eight enterprise customers runs an open-weight model in production [10], which would make the 29% a small cohort routing aggressively rather than a broad migration - one that stalls. Or policy decides it: a US export-control directive suspended access to Claude Fable 5 for the rest of June after it reached 22% of Opus 4.8 request volume in four days [6], and Chinese labs already take two-thirds of gateway video spend [18], so the cheap tier is one directive away from being unavailable rather than merely cheap.
What would prove the decoupling thesis wrong is simple enough to check: if Anthropic's high-stakes spend share slips below 72% while open-weight volume keeps climbing, then buyers are not protecting a tier, they are replacing one.
