Vercel AI Gateway index: open weight is 71.5% of tokens

Open-weight models now carry most of the AI traffic on Vercel's AI Gateway, and the money still goes somewhere else. In its October production index, published October 15, Vercel says open-weight models ran 71.5% of September tokens but took only about 23% of spend. The more useful number for an operator: nearly one in two teams changed their most-used model in a single month.

What actually happened

The data is anonymized, aggregate traffic through Vercel AI Gateway through September 2026. Spend is estimated at list prices, so real bills can differ. The headline findings:

• Open-weight share. 71.5% of tokens and about 23% of spend. Spend share was about 5% in June, so it more than quadrupled in three months.
• Cheap volume, small bill. DeepSeek-V4.1-Flash processed nearly half of September tokens and only 2.6% of spend.
• Anthropic's share fell. Anthropic's share of gateway spend dropped from 63.7% to 43.9%, under half for the first time. Frontier spend still grew tenfold since December. It is a bigger pie, not a smaller lab.
• Switching is normal. Nearly one in two teams changed their most-used model in September. Almost one in three changed their most-used lab, an index record.
• Teams run many models. Teams with 1B+ tokens used 42 models on average since December.
• Decision models. TypeSafe AI's Jev reached 35.3% of paid teams in its first two weeks, the fastest adoption in the gateway's history. TechCrunch reports Vercel itself uses Jev to classify harmful commands.

One caveat Vercel's own data cannot fix: large customers of OpenAI and Anthropic often call those APIs direct, so a gateway can undercount the big labs. TechCrunch makes the same point. Read this as "what multi-model teams do," not total market share.

Why model routing matters for your business

The pattern in this data is the one we build for: cheap models do the volume, expensive models do the hard parts. Classification, extraction, tagging and routing go to an open-weight or decision model at a fraction of a cent. Code review, contract summaries and the customer-facing answer go to a frontier model. The bill follows the hard parts, which is what you want.

Half of teams switching their main model in one month tells you something else: the model is not a long-term decision anymore. If a model name is hard-coded in 30 places in your app, every price change is a refactor. If it sits behind one routing layer with your own evals, it is a config change.

Three things to do this quarter:

• Tag every AI call by task. You cannot route what you cannot see.
• Price your top three tasks on two models each. One frontier, one open-weight. Use your own data, not a leaderboard.
• Keep an eval set per task. Twenty real examples with known good answers is enough to catch a bad swap.

Sources: Vercel, TechCrunch.