The new feature and its operating principle
On November 11, 2026, Google presented Gemini Decisions API, a service that evaluates up to 1,000 independent questions on the same set of information. The Gemini Decision 1 model does not generate natural-language text: it instead assigns probabilities to the options specified by the application. It is intended for repetitive operations such as classifying a request, setting a priority, or choosing an AI agent's next step, according to the source.
The mechanism is declarative: the software defines the permitted answers, the model evaluates them and returns a probabilistic judgment; rules, thresholds, and controls remain with the organization. As the source clarifies, a score does not amount to an authorization: detecting an anomaly in a case does not allow it to be rejected automatically. Google lists three permitted question types—a yes/no check, a choice among 2 to 255 alternatives, and an assessment on an ordered scale—which can be combined in the same request without resubmitting the documents.
Stated performance and terms of use
Google claims a median computation time of 50-100 milliseconds and an overall latency of 150-250 milliseconds, measured on contexts of up to 16,000 tokens and between one and 100 questions. The documented maximum limits—1,000 questions and 256,000 multimodal tokens—should not be automatically extended to those times. The source specifies that the shared context is counted only once, while instructions and criteria for individual questions add billable tokens.
The API is accessible via Gemini API in Google AI Studio, including for prototyping; at the time of the announcement, support in Google Cloud is described as forthcoming. Google cites internal applications, such as routing requests in Gemini Notebook with a latency reduction of up to 40% in some interactions, and detecting adversarial behavior in Search related to MCP tools. These are results reported by the company, not an independent comparison.
The comparison with other providers
OpenAI made its own Decisions API available on October 22, 2026, with a maximum of 400 questions per request and support for text, images, and JSON, context caching, and regional processing in the United States and the European Union. The source cites Datadog's use to assess the risk of shell commands produced by agents before execution. Microsoft-Decision-1 is in public preview in the Foundry catalog, accepts text and JSON, and derives from additional training of Qwen3.5-9B.
Cloudflare offers Clef and Clef-flash on Workers AI with Apache 2.0 weights; the hosted Clef-flash service has a context of about 24,000 tokens, while self-hosting reaches 256,000. Liquid AI offers d1-ultra via API and open-weight models announced on October 27 with the LFM3 family. Since November 4, Perplexity has introduced web grounding, with a reported increase from 51% to 86% correct answers on 100 questions of the PWS-Decisions benchmark: a limited test, not an independent verification, which increases time and costs.
What remains to be verified
The source emphasizes that the number of questions per request does not in itself indicate accuracy or speed under real-world conditions, and that price lists should be compared by individual API variant, not by family. The operational thresholds that separate automated cases from those to be forwarded to an operator must be tested on representative data. The actual regional availability of the services and the total costs when evaluating many questions on the same document also remain to be verified.
