Anthropic pulled in 65.1 percent of every dollar spent through Vercel’s AI Gateway in July while handling only 30 percent of the tokens that moved across it, according to Vercel’s own production index, reported by The Decoder on 18 August 2026. Divide one figure by the other and Anthropic charges 4.4 times what the gateway’s remaining providers, everyone routing traffic through Vercel besides Anthropic itself, charge on average per token. That is not a rounding gap. It is a sustained premium that developers keep choosing to pay.
Opus 4.8 led all spending on the gateway, with Fable 5 close behind at 13.2 percent, enough to put it in second place. Vercel’s data shows nine of every ten teams that adopted Fable 5 in July were new customers, which the company reads as fresh demand rather than existing users shuffling budgets. That reading cuts against a claim payments company Ramp made earlier this month, when it pointed to comparatively low Fable transaction volume as evidence the model was priced above what the market would bear. Vercel and Ramp are each looking at a different slice of the same funnel, spend routed through one inference gateway versus payment activity Ramp’s customers report, and neither slice is the whole market.
A 4.4x premium that survives a full quarter is a data point about switching costs, not just about pricing power. If Anthropic’s advantage were purely reputational, cheaper competitors with comparable benchmark scores would erode the gap month over month. Instead, Gateway token volume rose 59 percent and total spending rose 37 percent even as the average price per token across the whole gateway fell 13.6 percent, because customers who wanted to save money mostly did so by routing more volume to cheaper tiers, not by moving off Opus 4.8 for the workloads that matter most.
That split, cheap tokens for cheap tasks, premium tokens for the work developers will not risk on a second-tier model, is the mechanism behind the number. Code generation, agentic tool use and long-context reasoning are exactly the categories where a wrong answer costs an engineer more in debugging time than the token bill costs in dollars. Anthropic’s pricing behaves like it knows this. A provider can only sustain a 4.4x markup where the downside of switching (a broken agent loop, a regression a benchmark did not catch) outweighs the token savings on offer.
AI Insiders reported in July, from an earlier cut of this same Vercel index, that open-weight models had reached 29 percent of Gateway tokens while still capturing under 4 percent of spend. The pattern in the July data is the mirror image of that story: token share and dollar share are decoupling at both ends of the market, cheap open-weight capacity absorbing routine volume, and Anthropic’s frontier models absorbing the premium.
What would break this requires more than a rival matching Opus 4.8 on a leaderboard. It requires a competitor proving reliability on the specific agentic and coding workloads where Anthropic currently wins by default, sustained across enough teams that switching stops looking like a risk. Until that happens, expect the price gap to persist even as raw token costs keep falling.
One caveat matters here: this is one gateway’s traffic, weighted toward the web development teams that route inference through Vercel rather than a census of enterprise AI spending broadly, so the 4.4x figure describes that population, not the whole market.
Teams benchmarking Fable 5 or other challengers against Opus 4.8 should test on their actual agentic workloads before treating token price as the deciding factor, since Vercel’s own numbers suggest most developers are not.
The Decoder (Matthias Bastian, 18 August 2026), reporting on Vercel’s July AI Gateway production index.
