GRANITEAI / AI LEADERBOARDS
AI LeaderboardsFollow the token traffic.
Track which models gain share on Vercel AI Gateway. Compare daily usage trends and the split between open-weight and proprietary models.
Daily token share · Vercel AI Gateway · All modalities
On 2026-09-13, DeepSeek V4.1 Flash led named models with 68.0% of Vercel AI Gateway token traffic. Open weights accounted for 77.7%, closed weights 10.4%, named models awaiting classification 0.0%, and Vercel Other 11.9%.
100.0% of named-model token share classified. Vercel Other is excluded from this coverage figure.
Top model by token share Sep 13, 2026
DeepSeek V4.1 Flash68.0% of reported token trafficOpen weights
77.7%of reported token trafficClosed weights
10.4%0.0% remains unclassifiedToken share by model
See which models gain ground. Select a day to inspect the rankings, or switch to lines to follow each model over time.
| Model | Share |
|---|---|
| DeepSeek V4.1 Flash | 68.0% |
| Other | 11.9% |
| GPT 5.6 Luna | 4.1% |
| GLM 5.3 Flash | 3.0% |
| Step 3.7 Flash | 2.8% |
| DeepSeek V4 Flash 0731 | 2.5% |
| Muse Spark 1.3 Contributor | 2.0% |
| Gemini 3.8 Flash | 1.8% |
| Kimi K3 | 1.5% |
| Claude Opus 4.8 | 1.4% |
| Claude Opus 5 | 0.9% |
Open vs. closed weights
Track how token usage shifts between models with published weights and proprietary models.
Open weights
77.7%Closed weights
10.4%Unclassified
0.0%Vercel Other
11.9%Unclassified means named models awaiting evidence. Vercel Other is traffic grouped by the source without model names; it cannot be assigned to either weight class.
Check model classifications and sources
About the data
These are shares of token traffic through Vercel AI Gateway. They measure usage within that gateway, not global market share, user counts, or model quality. Vercel publishes percentages; absolute token counts are unavailable.
Each bar shows one day and adds up to 100%. Rankings and summary cards reflect the selected day, with the latest day shown first. The 7-, 30-, and 90-day controls change the history on the chart. They do not aggregate usage across the period.
Classifications refresh daily using exact model IDs and the Models.dev open_weights field, with publisher weight repositories checked when available. Conflicting or missing evidence stays unclassified. Each model lists its source and check date. Published weights can still carry license restrictions. The charts apply the current classification across the displayed history.
Unclassified means named models awaiting evidence. Vercel Other is traffic grouped by the source without model names; it cannot be assigned to either weight class.
AI Gateway Leaderboard Data © 2026 Vercel. Licensed under CC BY 4.0. Charts and weight groupings adapted by GraniteAI.