What Vercel AI Gateway data shows
Vercel published its September AI Gateway Production Index on September 17, 2026. The report uses anonymized, aggregate traffic routed through Vercel AI Gateway and covers data through August, with a separate look at early September model launches. Its headline finding is that open-weight models processed 56% of gateway token volume in August, up from 7% in December 2025. Those models represented 14% of estimated spend.
The report also says the average estimated price per token fell 23.2% in August, while the median team that exceeded ten million tokens in both months paid 7.6% less per token. Vercel notes that spend estimates use model providers’ published list prices, actual bills may differ, classifications can change, and earlier data may be revised.
Fact: model traffic moved quickly across providers
The index describes sharp changes in traffic and spend across model families. It reports that Google’s share of gateway token volume fell from 30% to 5% over the measured period, and that more than three quarters of the volume lost by Gemini 3 Flash moved to other labs. It also reports that GPT-6 Astra took 7.7% of gateway spend during its first twelve days, compared with 3.7% for Anthropic’s Fable 5.1 over each model’s first twelve days.
These are observations from one gateway, not a census of the AI market. They do not prove that one model is universally better, that open-weight deployments are cheaper in every system, or that a team should copy the aggregate. Vercel explicitly says its movement analysis compares changes among the same teams and does not trace individual tokens from one model to another.
Seuil analysis: model choice is becoming product infrastructure
The useful signal is not which logo won the month. It is how quickly production workloads can move when price, capability, or reliability changes. A product that hard-codes one model into every workflow turns a market change into a rewrite. A product that separates task policy from provider integration can retest and move deliberately.
That separation has three layers. First, define the task: extraction, classification, drafting, tool use, vision, or long-horizon reasoning. Second, define the acceptance criteria for that task. Third, route only to models and providers that meet the criteria. The same user experience may use a small model for intent detection, deterministic code for permissions, a stronger model for a difficult transformation, and a human for a consequential approval.
Do not turn routing into random model switching
Multi-model architecture adds failure modes. Outputs may differ in structure, refusal behavior, tool syntax, context limits, and latency. A fallback can return a technically successful answer that violates the product contract. The product therefore needs a normalized interface, typed outputs where appropriate, model-specific tests, and metadata that records which route served each request.
Vercel’s AI Gateway documentation says its unified API supports model and provider changes without rewriting application sections. Its model fallback documentation explains that backup models are tried in order after provider routes for a model fail. That is useful availability machinery. It is not a quality guarantee. Teams still need evaluation cases for every permitted route and a safe response when no route meets the product threshold.
Measure cost per successful task
The index’s price trend will tempt teams to optimize for cost per token. That metric is incomplete. A cheaper model may require more retries, larger prompts, more human review, or a second model to repair its output. A premium model may be wasteful on simple classification but economical on a task where one correct pass prevents manual rework.
The operational metric should connect cost to a completed, acceptable task. Track input and output tokens, cache use, latency, retries, fallback rate, evaluation score, and human-review time. Then compare routes on the same examples. Vercel’s current budget documentation provides team, project, API key, and user spend caps, while noting that the request that crosses a limit can still complete. Spend controls limit damage. They do not decide which workload deserves which model.
Preserve a consistent user contract
Users should not need to understand the routing table. They should experience consistent controls, response structure, citations, error handling, and recovery. If a fallback cannot support an attachment, a tool, or a required source trace, the interface should say so rather than silently returning a weaker answer. Product design turns infrastructure variance into a stable contract.
This is also where latency matters. Provider ordering and timeouts can improve resilience, but an aggressive timeout may abandon a model that would have produced the best answer, and some timed-out requests may still incur charges. Route policies should reflect user expectations: a short wait for autocomplete, a longer wait for a researched analysis, and explicit progress for an agent completing several steps.
A four-part response for product teams
Teams that do this can benefit from a fast-moving model market without making their product feel unstable. Teams that do not will accumulate a different kind of lock-in: undocumented prompts, implicit quality assumptions, and interfaces built around one provider’s quirks.
- Inventory tasks. List each AI-assisted task, its risk, volume, latency target, and required capabilities.
- Build route-specific evals. Use representative examples and failure cases before approving a model or provider.
- Log the decision. Record model, provider, version, latency, cost, fallback, and user-visible outcome.
- Review monthly. Retest when pricing, models, prompts, tools, or providers change. Do not migrate because a leaderboard moved.
The decision is architectural, but the outcome is product quality
Vercel’s report is a snapshot of its own gateway traffic, and its exact shares will change. The durable conclusion is narrower: production teams are already distributing workloads across model profiles, and cost alone does not explain the movement. Product leaders should treat model selection, routing, evaluation, and spend controls as one product system.
If your AI feature depends on one provider today, the next useful exercise is not an immediate migration. It is a controlled test: choose one high-volume or high-risk task, define the pass criteria, compare two routes, and inspect the user-visible difference. Seuil Digital can help frame that product and experience decision before it becomes an emergency rewrite.