Google launched its Gemini Selections API on Nov. 11, 2026, letting software program ask an AI mannequin for a bounded judgment with out producing a written reply. The endpoint returns possibilities and glued selections that code can use on to route requests, test insurance policies or rating content material in opposition to a rubric.
Powered by Gemini Choice 1, the service helps sure/no questions, selections from a provided record and scores in opposition to predefined ranking scales. It’s accessible by means of the Gemini API in Google AI Studio, in keeping with Google’s launch announcement.
The sensible change is within the response software program receives: a price it could possibly department on, accompanied by confidence, with out having to parse a chatbot’s prose. Generated explanations, device calls with arguments and multi-step reasoning stay makes use of for Google’s Interactions API, its documentation says.
Google lists a worth of $0.10 per million enter tokens and no output-token cost. Shared enter is billed as soon as inside every request, with every query including its personal instruction and standards tokens. OpenRouter’s itemizing reveals the identical enter and output pricing.
A request can mix textual content, pictures, audio, video and PDFs with as much as 1,000 unbiased questions. The full context restrict is 262,144 tokens throughout shared enter and all questions. The questions can’t see each other’s solutions, so code should mix their outcomes.
How can code use confidence?
Illustrative code path after validation on related examples. A handed confidence threshold isn’t a security assure.
Illustrative code path for an ambiguous outcome: use a fallback or human evaluate. Deal with out-of-scope enter explicitly.
The mannequin provides judgments; software program owns the motion and escalation guidelines.
Confidence should be validated on related labeled information. Alternative can choose an choice even when none suits; scope checks or an different choice stay mandatory. Passing a threshold doesn’t assure security; Google recommends human or second checks for high-stakes actions. Traces don’t encode latency or quantity.
Google stories end-to-end latency of 150–250 milliseconds for contexts as much as 16K tokens, with latency almost flat throughout 1–100 questions per request. These reported measurements cowl a smaller workload than the utmost context and query limits.
Google says Gemini Pocket book already makes use of the service to decide on mannequin tiers and pondering ranges, with as much as 40% decrease latency for on a regular basis queries whereas reserving deeper reasoning for advanced requests. It additionally says Search makes use of the API to detect adversarial conduct and content material as a security layer for its instruments.
Gemini Selections confidence nonetheless wants testing
Every alternative or ranking comes with a separate confidence estimate of whether or not the highest reply is right. Google’s documentation recommends setting thresholds with labeled examples from the developer’s personal visitors, weighing the prices of false positives in opposition to false negatives.
A set record additionally creates a particular failure case: the mannequin distributes chance among the many provided choices even when none suits. Google recommends a confidence threshold, a separate test that the enter belongs throughout the query’s scope, or an explicitly described catch-all choice.
Consumer-supplied textual content can argue for its personal classification, the documentation warns. Google recommends specific standards, adversarial assessments, and a human or second test for high-stakes actions akin to payouts.
