Gemini Decision 1 lands on Vercel AI Gateway
Vercel's gateway now serves Google's multimodal decision model, which answers with probabilities, choices and scores instead of generated text.
Gemini Decision 1 arrives on Vercel's AI Gateway
On 11 November 2026, Vercel added Gemini Decision 1 from Google to its AI Gateway, the layer it runs for model access, usage tracking and cost. The model is not a text generator in the usual sense. It is a multimodal decision model that reads text, images, audio, video and PDF documents and answers with Boolean probabilities, choices and scores.
The changelog entry on Vercel's site, credited to Aayush Kapoor, Jerilyn Zheng and Zachary Chen, names the jobs it is meant for: routing requests, checking policies, filtering agent tools and ranking results.
How the decision calls work
Builders reach the model as google/gemini-decision-1 through the AI SDK, a TypeSafe-compatible API, or a dedicated Decisions API. In the AI SDK, the entry point is an experimental decide function imported from ai. A call takes a model id, a piece of state such as a customer message, and a set of named questions. Each question carries a type, such as boolean, plus an instruction describing what should be evaluated. The result comes back as an answers object keyed by question name, which the sample code prints.
Input support depends on the route. AI SDK 7 accepts PDF, audio and video in addition to text and images. The TypeSafe-compatible and Decisions APIs take text and images only.
Because the model returns a decision rather than generated prose, no output tokens are produced. Gateway accounting still applies: decision calls appear in logs and count toward budgets alongside other model requests. Vercel points readers to its decision documentation for the Choice, Score and Boolean question types, and to its model catalogue and decision model guides for choosing one.
What changes for people building apps
The practical shift is architectural. A lot of product logic that teams currently write as prompt-and-parse loops, where a chat model is asked to reply with a JSON verdict and the code then validates it, can move to a model whose output is already structured. Routing an incoming request to the right handler, screening user content against a policy, deciding which tools an agent may call, or ordering search results are all classification tasks. Getting a typed answer directly removes a parsing step and the failure modes that come with it.
It also changes cost shape. Vercel's framing is that decision calls avoid output tokens entirely, so a high-volume classification pass is not billed like a generation pass. The source does not give prices, so the actual saving is unknown.
Limits and open questions
The announcement is short and leaves several things unstated: no pricing, no rate limits, no latency figures, and no context or input size limits for the multimodal inputs. It does not say whether the model is generally available or in preview, or whether Vercel hosts it or passes it through. The experimental_ prefix on the AI SDK function suggests that surface may still change. Audio, video and PDF support is tied to AI SDK 7, so older SDK versions cannot use those inputs. How well the returned probabilities are calibrated, and whether score questions behave consistently across modalities, is not addressed.
For building apps with AI
The wider signal is that decision models are becoming a separate category from chat models, with their own APIs and their own cost model. If you are assembling an app that classifies or routes with AI, that logic is now a distinct piece you can plan for rather than an improvised prompt. See Draftship's notes on adding AI to your app for where that kind of call sits in a build. Draftship has no stated support for Gemini Decision 1 or AI Gateway.
Primary sources
- Gemini Decision 1 now available on AI Gateway · Vercel, Nov 11, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.