Anthropic's share of spending through Vercel AI Gateway fell from 63.7% to 43.9% in September 2026, dropping below half for the first time. Vercel published the figures in its October index on October 15. Open-weight models handled 71.5% of tokens and about 23% of spending over the period. This describes activity routed through Vercel's service, rather than the companies' global market shares.
Open-weight spending share rose from roughly 5% in June, more than quadrupling. The gap between token share and spending share is central: usage volume and money are different measures. A model can process many tokens without taking a similar proportion of the budget, so those figures should not be read as measuring the same thing.
DeepSeek leads volume; Opus stands out in spending
DeepSeek alone handled 58% of all Gateway tokens, with about 6.6% of spending. This refers to the company's models, rather than one individual model. Claude Opus 5.5, meanwhile, led spending share among models in September's launch table, at 11.03% during its launch window. The unit of comparison and each table's measurement period both matter.
Anthropic's declining share describes a relative change within Gateway spending. It does not by itself prove that absolute spending on the company fell, and does not measure revenue from all its customers. Similarly, token leadership is not a quality ranking. The index shows how usage and budgets are distributed within the service, but cannot answer which model fits every task.
More teams are changing their most-used model
Vercel says 48.0% of teams included in its switching measure changed their most-used model in September, while 28.6% changed their most-used AI lab. Both were index records. Replacing a team's leading model does not necessarily mean abandoning the previous model entirely or working exclusively with a new provider.
In the scorecard covering the first two weeks of models launched in September, Claude Sonnet 5.5 scored 87.8. It was followed by Gemini 3.8 Flash at 81.3, Claude Opus 5.5 at 79.9, GPT-6 Luna at 79.4 and DeepSeek-V4.1-Flash at 78.4. These are Vercel scorecard figures, not an independent quality test for every task or a measure of answer correctness.
For development teams, the figures offer a reason to separate three questions: how much a model is used, how much is spent on it and how well it fits their work. Spending share can change meaningfully while a model remains important to its users. The index supplies a focused view of September on Gateway; a migration decision requires testing the team's actual tasks and costs.