Models#Model inference

Google ships Gemini Decision 1: a model that scores answers instead of writing them

Google's Decisions API and Gemini Decision 1 launched Nov. 11: one forward pass scores candidate answers at $0.10 per million input tokens, zero output cost.

Google blog header image reading Gemini Decisions API

Where the rest of the industry races on generation, Google’s November 11 release goes the other way. Gemini Decision 1 does not write answers: it scores candidate options in a single forward pass and returns each result with a confidence value. Google CEO Sundar Pichai announced the model on X, saying it is live in the Gemini API in Google AI Studio alongside the new Decisions API. Chinese-language coverage followed the same day — InfoQ’s AI 前线 and the commentator 宝玉 (dotey) both walked through the design.

Key points

  • One pass, no generation: a single input can mix text, images, audio, video and PDFs and carry up to 1,000 questions. Three question types are supported — yes/no, multiple choice with 2 to 255 options, and rubric-based scoring. Google lists 150–250 ms end-to-end latency, and says adding questions barely changes it.
  • Input-priced, output-free: the API bills $0.10 per million input tokens and charges nothing for output. Google says that across a shared benchmark suite it is nearly 5x cheaper than generating answers with Gemini 3.5 Flash-Lite through the Interactions API and parsing them.
  • The benchmarks are Google’s own: the company reports leading results with the lowest calibration error on benchmarks covering agent routing, reasoning, ranking and multimodal understanding; InfoQ’s report phrases it as first place in 7 of 9 tests. None of this has independent replication yet — test on your own tasks before switching.

Google blog header image reading Gemini Decisions API

Why a decision model

Google positions the API for fast, calibrated multimodal decisions: routing a support ticket, checking content policy, choosing which model should handle a request. Those calls want a judgment with a confidence score, not a paragraph. The traditional pattern — generate text, parse the answer — is slow, can fail to parse, and pays for output tokens. The Decisions API collapses all three costs into one call that returns structured results with zero output billing.

The calibration claim is the part worth watching. Google says the returned probabilities are calibrated, meaning that when the model says 90 percent it is right about nine times in ten — so product code can set thresholds directly, auto-executing above 0.7 and escalating below it. Reliability of judgment is being sold as part of the API, which is new.

The blog’s own example makes the shape concrete: upload one photo of an office kitchen counter and ask three questions at once — whether liquid has been spilled, how it should be handled, and a 0-to-2 mess score. Each question returns its own answer and confidence. That one-context-many-small-questions shape is what routing, moderation and quality-control calls look like in production, and it is where this model and a generative model divide the work.

Analysis

  • The cost structure favors high-volume classification: with no output billing, routing and moderation calls get cheap fast, and the bill scales with input volume — the more questions ride on one shared context, the thinner the per-question cost.
  • Treat “first place” as a vendor claim: the benchmark numbers and the 5x figure come from Google and secondary reports, as with Gemini 4 Carbon’s launch; wait for independent evaluation before they reach a selection decision.
  • A decision model does not replace generation: it only chooses among provided options. Anything outside the candidate set still needs a generative model — the two divide the work rather than compete.

Summary

If your product makes many “pick one”, “who should handle this” or “is this allowed” calls, the Decisions API is worth a trial round in AI Studio on real tasks: record latency, confidence distributions and the bill, then decide whether those calls migrate off generative models. The benchmark lead is Google’s claim; the migration case has to be yours. Prove the decision-type calls on their own first, then consider splitting the product between them and a generative model: candidates inside the set go to Decisions, everything outside stays generative.