Google’s new AI can make 100 decisions in under a second, without writing an answer

Google introduced its Gemini Decisions API on Nov. 11, 2026, letting software ask an AI model for a bounded judgment without generating a written answer. The endpoint returns probabilities and fixed choices that code can use directly to route requests, check policies or score content against a rubric.

Powered by Gemini Decision 1, the service supports yes/no questions, choices from a supplied list and scores against predefined rating scales. It is available through the Gemini API in Google AI Studio, according to Google's launch announcement.

The practical change is in the response software receives: a value it can branch on, accompanied by confidence, without having to parse a chatbot's prose. Generated explanations, tool calls with arguments and multi-step reasoning remain uses for Google's Interactions API, its documentation says.

Google lists a price of $0.10 per million input tokens and no output-token charge. Shared input is billed once within each request, with each question adding its own instruction and criteria tokens. OpenRouter's listing shows the same input and output pricing.

A request can combine text, images, audio, video and PDFs with up to 1,000 independent questions. The total context limit is 262,144 tokens across shared input and all questions. The questions cannot see one another's answers, so code must combine their results.

How can code use confidence?

Illustrative code path after validation on relevant examples. A passed confidence threshold is not a safety guarantee.

The model supplies judgments; software owns the action and escalation rules.

Confidence must be validated on relevant labeled data. Choice can select an option even when none fits; scope checks or an other option remain necessary. Passing a threshold does not guarantee safety; Google recommends human or second checks for high-stakes actions. Lines do not encode latency or volume.

Sources and evidence

Sources

1. Google's November 11 launch announcement
2. Google Decisions API documentation

Supporting facts

• The Decisions API returns typed yes/no answers, fixed-list choices and ordered-rubric scores with probabilities and confidence, generating zero text tokens. [1] [2]
• The Decisions API accepts shared text, image, audio, video and PDF input and 1 to 1,000 independent questions within a 262,144-token context window covering both shared input and all questions. [2]
• The API documentation defines confidence as an estimated probability that the top answer is correct and recommends setting thresholds using labeled examples from the developer's own traffic and the costs of false positives and false negatives. [2]
• Choice questions allocate probability among the supplied options even when none fits; Google's documentation recommends confidence thresholds, a separate scope check or an explicitly described catch-all option. [2]
• Google's documentation warns that user input can argue for its own classification and recommends explicit criteria, adversarial testing and a human or second check for high-stakes actions such as payouts. [2]
• The documentation positions the Decisions API for routing, policy checks and ranking, while directing generated explanations, tool calls with arguments and multi-step reasoning to the Interactions API. [2]

Method

Qualitative schematic of shared input, independent question types and typed responses. The selectable panel illustrates confidence-threshold handling in application code, with no numerical threshold or observed traffic assumed.

Scope

Confidence must be validated on relevant labeled data. Choice can select an option even when none fits; scope checks or an other option remain necessary. Passing a threshold does not guarantee safety; Google recommends human or second checks for high-stakes actions. Lines do not encode latency or volume.

Infographic by CryptoSlate

Google reports end-to-end latency of 150–250 milliseconds for contexts up to 16K tokens, with latency nearly flat across 1–100 questions per request. Those reported measurements cover a smaller workload than the maximum context and question limits.

Google says Gemini Notebook already uses the service to choose model tiers and thinking levels, with up to 40% lower latency for everyday queries while reserving deeper reasoning for complex requests. It also says Search uses the API to detect adversarial behavior and content as a safety layer for its tools.

Gemini Decisions confidence still needs testing

Each choice or rating comes with a separate confidence estimate of whether the top answer is correct. Google's documentation recommends setting thresholds with labeled examples from the developer's own traffic, weighing the costs of false positives against false negatives.

A fixed list also creates a specific failure case: the model distributes probability among the supplied options even when none fits. Google recommends a confidence threshold, a separate check that the input belongs within the question's scope, or an explicitly described catch-all option.

User-supplied text can argue for its own classification, the documentation warns. Google recommends explicit criteria, adversarial tests, and a human or second check for high-stakes actions such as payouts.