Skip to the cards
Filed
Source
David Jones
Loans
347

Google Gemini Decisions API Evaluates 1,000 Questions per Request

Gemini Decisions API scores options without generating text, with multimodal inputs, 150–250 ms test latency and a $0.10-per-million-token price.

Google Gemini Decisions API Evaluates 1,000 Questions per Request

Google’s Gemini Decisions API is designed to make a common AI workflow faster: evaluating a set of conditions or choosing among defined options without generating a text response for every question. Developers can submit as many as 1,000 questions in one request, and Google says the service can return results with 150 to 250 milliseconds of end-to-end latency in tests using contexts of up to 16,000 tokens.

That makes the API a specialized decision layer for applications that need to classify, route or score inputs. Its value will depend on whether those decisions are reliable in the data and operating conditions a business actually uses. A confidence score can help determine when to automate a step and when to send a case for review, but it does not establish that the underlying judgment is correct.

Scoring options instead of generating answers

The API is powered by Gemini Decision 1. Traditional generative models produce a response token by token; Google’s decision model instead assigns probabilities to options specified by the developer. An application can use it to check whether a condition is met, select from a predefined set, or score an input against criteria in a rubric.

This approach may suit tasks where the application needs a structured result rather than prose. For example, a developer could define the choices a routing workflow should consider or provide criteria for judging whether an input meets a requirement. The model’s output can then inform the next step in the application. The API evaluates the choices and criteria supplied to it; teams still need to decide what those options mean and what action should follow each result.

Google says the API accepts images, audio and video as well as text, and supports a context window of up to 256,000 tokens. That gives developers room to evaluate multimodal material alongside longer context. It also means teams should assess performance and reliability using the kinds of inputs their own workflows will send, rather than assuming a result observed with short text prompts will carry over to every format.

In Google’s tests, latency stayed relatively stable as the number of questions rose from one to 100. The reported 150 to 250 milliseconds applies to tests with contexts of up to 16,000 tokens; Google’s information does not establish that the same latency applies to every context length or to a request containing the maximum 1,000 questions. Teams planning a high-volume workflow will need to measure response times with their own request sizes and input types.

Pricing and current availability

Google lists the price at $0.10 per million input tokens, with no additional charge for output tokens. Shared input is billed once per request, regardless of how many questions are evaluated, while each question adds its own instruction and criteria tokens. As a result, bundling related checks into one request could affect the cost of a workflow, although the actual bill will depend on the amount of shared context and the instructions attached to each question.

The API is available through Google AI Studio, and integrations with developer frameworks are rolling out. Google says support for Gemini on Google Cloud is expected later but has not provided a release date. Organizations considering it for production should account for that availability when choosing where to prototype and where to run deployed workloads.

Google is already applying the technology in Gemini Notebook to route queries to different AI models according to request complexity. The company says this lowered latency for everyday queries by as much as 40%. It also uses the API in Search to identify adversarial behavior involving Model Context Protocol (MCP) tools. These examples show two different roles for a decision model: selecting a processing path and assessing whether an input or interaction warrants a response.

How it compares with OpenAI’s offering

OpenAI’s Decisions API became generally available in October and supports up to 400 questions per request, compared with Google’s limit of 1,000. Both companies list a standard price of $0.10 per million input tokens and charge no output-token fees. OpenAI also offers discounted pricing for cached inputs, which may matter to applications that repeatedly process the same context.

Google supports a wider range of input formats than OpenAI’s API, including audio and video. That difference may be relevant to teams building workflows around mixed media, while the question limit and pricing structure matter to workloads that batch many checks or reuse context. A feature comparison alone won’t show which service fits better: developers also need to consider integration with their existing tools and how each system performs on their own tasks.

Lian Jye Su, chief analyst at Omdia, sees possible uses beyond model routing, including policy and risk assessment and credit scoring. He described the model as suited to producing decisions from unstructured, changing inputs. Su’s view is that the key comparison is how well each service fits an enterprise’s development environment and how it behaves in real operating conditions. He argued that decision quality and fewer misjudgments matter more than published benchmarks alone.

What teams should validate before deployment

Su warned that changing data or deliberately manipulated inputs can lead to incorrect judgments. A confidence score may indicate how strongly a model favors an option, but it cannot by itself tell a team whether that option is appropriate. Teams should define what happens when the model is uncertain, test cases that fall near decision boundaries, and review whether the result is supported by the context available to the application.

Su recommends connecting decision models to databases with historical transactions and established business policies. That gives an application relevant organizational information to draw on, rather than asking the model to assess a request from its immediate input alone. The quality and freshness of that context will affect the decision process, so teams need to consider which records and rules a workflow can access.

Enterprise adoption also depends on how Google handles data retention and sovereignty, Su said, and whether those requirements can be met while keeping the service’s cost and performance advantages. Separately, Tulika Sheel, senior vice president at Kadence International, highlighted accountability as AI agents take on more autonomous work. Organizations will need to make clear who is responsible for decisions and actions taken on a user’s behalf.

For administrators and development teams, the practical next step is to test the API against representative inputs, measure latency and costs at the request sizes they expect, and set a review path for uncertain or consequential outcomes. The headline limit of 1,000 questions offers room to batch checks, but production suitability will come down to the quality of the decisions, the context behind them and the controls around what happens next.

Source: David Jones · www.infoworld.com

Filed alongside