Vercel published its November AI Gateway Production Index on 9 November 2026, reporting decision-model traffic and frequent model switching. Its anonymized, aggregate October observations raise questions about research workload measurement and experiment records; they describe activity on one provider’s gateway.
Decision-model requests and spending measure different things
Vercel reports that decision models handled 26% of October gateway requests but accounted for 0.53% of estimated spending. Its spending estimates use published list prices, so actual bills can differ.
The company’s decision-model documentation describes structured choices, scores and Boolean probabilities. These outputs suit classification and routing. For a university assessing research software, the distinction is between the kind of work performed and the volume of traffic: a request count alone does not establish either scientific usefulness or budget impact.
The switching finding has a specific denominator
About half of the report’s switching cohort changed their most-used model in October. That cohort comprises paid teams processing at least 10 million tokens in both September and October; most-used is measured by tokens. Vercel measures each model version’s age from its first gateway availability.
For researchers, switching raises a provenance question: which model actually produced an output? Vercel’s fallback documentation says configured backups can run when a primary model fails or is unavailable, and routing metadata identifies model and provider attempts. A configured model name and the model that answered can therefore differ. That distinction matters when documenting a reproducible AI experiment.
What the index means for universities and research offices
The institutional implication is to separate adoption, execution and evaluation. Adoption concerns which services teams use. Execution concerns the route taken by a particular request. Evaluation concerns whether the resulting answer or action meets the research task’s requirements. The index supplies usage observations; task-specific evidence belongs in a model evaluation suite.
Vercel’s decision documentation calls for testing representative cases with agreed outcomes and measuring incorrect application actions and human review workload before automatic execution. Its request-log documentation describes served-model and provider fields, routing details and exports of loaded request rows. Those are concrete records an institution can distinguish from an aggregate adoption trend.
For research offices considering how to govern language models for research, our reading is that the report strengthens the case for recording changes and evaluating their consequences. It establishes no universal replacement interval. Vercel’s observability documentation also treats requests, tokens, spending and latency as separate metrics, with views at team and project level. Preserving those distinctions makes the November findings more useful than treating model churn as a quality ranking.
Sources
- Vercel: AI Gateway Production Index, November 2026, published 9 November 2026. Checked 2026-11-14.
- Vercel: Decision. Checked 2026-11-14.
- Vercel: AI Gateway Model Fallbacks. Checked 2026-11-14.
- Vercel: AI Gateway Request Logs. Checked 2026-11-14.
- Vercel: AI Gateway Observability. Checked 2026-11-14.
