The report also says open-weight models’ share of spend has more than quadrupled since June. Open-weight models are moving from token-volume dominance toward a meaningful, though still minority, share of production inference spend.
The report attributes token growth largely to cheaper models and spend growth to more expensive models, with Kimi K3 contributing nearly half the increase in open-weight spending and DeepSeek-V4.1-Flash processing high volume at a low cost per token. September recorded the largest monthly gains in open-weight token share and spend share in the index’s history, after open-weight models first processed a majority of gateway tokens in August. These are Vercel AI Gateway observations, not a census of enterprise AI usage or the entire model market.
Teams should separate model selection by workload volume from budget impact; the highest-volume model may account for only a small share of spend. Whether open-weight models continue gaining spend share in subsequent index editions, rather than only token volume, will show whether this shift persists. Vercel also reports that frontier-model spend grew tenfold since December, so the evidence points to expanding use across both categories, not a simple replacement of frontier models.