The short version
Cheap open-weight models now do most of the work in production AI apps, while expensive frontier models still get most of the budget. Vercel’s gateway data for September 2026 shows open-weight models processing 71.5% of tokens for 23% of spend, and almost one in three teams changing their main AI lab in a single month. If you are adding generative AI to an existing product, plan for several models behind one interface from the start.
What did Vercel’s AI Gateway index find?
Vercel’s AI Gateway is a single API that lets applications call models from many providers. Its October 2026 Production Index, published on October 15, summarizes anonymized traffic from December 2025 through September 2026. The headline result: open-weight models processed 71.5% of all tokens in September, and their share of estimated spend more than quadrupled since June to 23%. The Register reported the same figures on October 17.
The money still sits at the top. Spend on frontier models grew roughly tenfold over the period, according to Vercel, and Anthropic’s Claude Opus 5.5 took 11.03% of all gateway spend in its first two weeks, more than twice the share of any other September launch. Anthropic’s overall share of spend fell below 50% for the first time, from 63.7% to 43.9%. Meanwhile open-weight spend grew more than seventyfold, and the average price per token dropped 47% as teams pushed high-volume work onto cheaper models.
Vercel also found that 68% of spend on Chinese-developed models now runs through US-based inference providers, up from under 10% in December. In practice, teams are adopting models such as DeepSeek, Kimi and GLM while keeping the hosting in the US.
Why did a decision model become the most adopted?
The fastest-adopted model in the gateway’s history does not write text at all. TypeSafe AI’s Jev, released in limited access on September 15, takes program state and a typed question and returns a decision with a probability: a yes or no, a choice from a fixed set, or a score. Vercel says it reached 35.3% of paid teams within two weeks, yet accounted for only 0.27% of spend. TechCrunch, which reported the adoption figure on October 16, notes that Vercel itself uses Jev instead of a frontier model to classify harmful commands.
That pattern explains the wider numbers. Much of what AI features do in production is not open-ended writing but small decisions: route this ticket, flag this transaction, retry or stop this agent step. Those calls are frequent, need low latency and work better with a fixed output schema. The gateway data suggests teams are moving that layer to cheap, specialized models and keeping frontier models for reasoning-heavy steps.
How often are teams switching models?
Very often. In September nearly half of teams changed their most-used model, and almost one in three changed their most-used lab, a record for the index. Teams processing at least one billion tokens used an average of 42 different models. For anyone building on Vercel or a similar stack, the model is no longer a fixed architectural decision. It is a configuration value that will change several times a year.
What are the limits of this data?
The index reflects one gateway’s customers, not the whole market. TechCrunch points out that Vercel’s view is necessarily limited and may undercount large labs whose customers call proprietary APIs directly. Vercel estimates spend at published list prices, so negotiated discounts are not reflected, and its definition of open-weight is broader than in earlier reports. Treat the numbers as a strong signal about how web and product teams build, not as market share.
What it means for US & EU software teams
The main lesson is architectural. Applications that call one provider’s SDK directly from business logic will be expensive to change, and the data shows teams do change. A thin model layer, whether a gateway or your own abstraction, lets you swap models per task, run A/B evaluations and fail over during outages without touching product code.
The second lesson is cost. Splitting work by task type, with small or open-weight models for classification, extraction and routing, and frontier models only where quality clearly pays off, is how the teams in this dataset cut the average price per token by almost half. That only works if you measure quality per task with your own evaluation set rather than trusting public benchmarks.
For EU teams and regulated US sectors, the routing layer is also where compliance lives. Every model and inference provider is a data processor. Before moving traffic to an open-weight model on a new host, check where it runs, whether prompts are retained and whether it appears in your data processing records. A model that is cheaper but hosted outside your approved regions can create GDPR exposure that outweighs the savings.
What engineering teams should do this quarter
- Inventory model calls. List every place your product calls an LLM, which model it uses and what the call actually does.
- Separate decisions from generation. Identify calls that return a label, score or yes/no and test a small or specialized model for them.
- Add a model layer. Route calls through a gateway or internal interface so models can change by configuration, not by code release.
- Build a task-level eval set. Keep a few hundred real examples per task to compare models before any switch.
- Review providers for data residency. Record where each model is hosted and whether prompts are stored before routing production data to it.
Frequently asked questions
What share of AI traffic do open-weight models handle?
On Vercel’s AI Gateway, open-weight models processed 71.5% of tokens in September 2026 and accounted for 23% of estimated spend, according to the October 2026 Production Index published on October 15, 2026. The figures cover Vercel’s gateway customers, not the entire AI market.
What is Jev and why was it adopted so fast?
Jev is a decision model from TypeSafe AI, released in limited access on September 15, 2026. Instead of generating text, it returns typed answers such as yes or no, a choice from a fixed list, or a score, with a probability. Vercel says it reached 35.3% of paid teams in two weeks because it handles frequent classification and routing calls quickly and cheaply.
Should we switch our product to open-weight models?
Not wholesale. The data shows teams mixing models by task: cheap open-weight or specialized models for high-volume, simple calls and frontier models for complex reasoning. Test each task against your own evaluation set and check hosting location and data retention before moving production traffic.
Why does model switching matter for software architecture?
In September 2026 nearly half of Vercel’s gateway teams changed their most-used model and almost one in three changed their main lab. If model calls are hard-coded to one provider, each change becomes a code release. A routing layer turns it into a configuration change and enables per-task evaluation and failover.
Sources
Vercel — AI Gateway Production Index: Open weight takes 71.5% of tokens and 23% of spend (October 15, 2026)
TechCrunch — Jev is Vercel’s fastest-adopted AI model ever (October 16, 2026)
The Register — Open-weight models win hearts, minds, and wallets (October 17, 2026)
Vercel — Jev is the fastest-adopted model in AI Gateway history (September 18, 2026)