Glonce

Google launches Gemini Decisions API for fast closed-set choices

toolsThe Glonce editorsHacker News🔥82Glonce rating82Impact62Novelty72Relevance100Surprise50Players88Credibility85
Google launches Gemini Decisions API for fast closed-set choices

Google has introduced the Gemini Decisions API, a new endpoint in the Gemini API (available in Google AI Studio) for fast, closed-set decisions. It runs on a new model, Gemini Decision 1 (gemini-decision-1). The endpoint is meant for bounded choices such as tool selection, policy checks, model routing and context grading.

The key difference from a normal language model call is that Gemini Decision 1 does not generate answer text token by token. It scores the candidate options you supply in a single forward pass and returns probabilities and confidence scores. Google gives 150-250 ms end-to-end latency, of which 50-100 ms (median) is model compute. Those figures are described for multimodal inputs and contexts up to 16K tokens, and latency stays nearly flat from 1 to 100 questions per request.

One request takes a single shared multimodal input (text, images, audio, video and PDFs, up to 256K tokens) plus up to 1,000 independent questions. Examples Google gives are a call recording paired with account metadata, or a scanned PDF claim with damage photos. Resolution on image and video inputs can be tuned to control visual detail and token cost. There are three question types, and they can be mixed in one request:

  • noul (Yes/No): estimates the probability that a yes/no condition holds, with optional descriptions of the true and false criteria.
  • choice (Multiple Choice): selects from 2 to 255 choices and returns the winner plus normalized probabilities for every candidate.
  • score (Ordered Rubric): grades the input against an ordered rubric starting at 0 and returns a continuous weighted score and per-level probabilities.

Every answer carries a confidence score from 0 to 1, so application code can send ambiguous cases to a fallback or a human reviewer. The input field uses the same Content block format as the Interactions API. Google's Python example calls client.decisions.create with model gemini-decision-1, passes a photo of a breakroom counter and a text note, and asks three questions at once: whether there is an uncleaned spill (noul), which cleanup action fits (choice, with wipe_with_cloth, dispatch_janitorial or ready_for_meeting), and how severe the mess is (score, over a three-level rubric).

On performance, Google says that on benchmarks covering agent routing, reasoning, ranking and multimodal understanding, Gemini Decision 1 leads across the suite and has the lowest calibration error. The post gives no benchmark names or values. On cost, Google says Gemini Decision 1 with the Decisions API is nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API, helped by zero output tokens being billed.

Pricing is $0.10 per 1M input tokens and $0 for output tokens, with an API no-cost tier in Google AI Studio for prototyping. The shared input is billed once per request however many questions are asked; each question adds only its own instruction and criteria tokens. There are no output-token charges, variable thinking-token bills, or separate audio transcription and OCR costs.

Google says teams across Google already use the API as a low-latency control plane. Gemini Notebook uses it as a real-time router that classifies intent in milliseconds and picks the model tier and thinking level, serving everyday queries with up to 40% lower latency while keeping deep reasoning models for complex tasks. A new rapid canvas experiment in Stitch uses it to generate playable wireframes of a product flow in seconds and to apply voice or tap edits almost instantly. CC from Google Labs uses it to pick an emoji reaction from a palette of more than 200 options. Google Search uses it as a safety layer for MCP tools, to detect adversarial behavior and content; the post says only that it shows high efficacy within the latency envelope on adversarial evals.

To start, developers can get an API key in Google AI Studio, try the Doorstep Parcels, Speed Sketch and Find the Zoo Keeper demo apps, and follow the Python, JavaScript and REST guides. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode alongside the google-genai SDKs, with support coming soon to Gemini in Google Cloud.

Key facts

  • Gemini Decision 1 (gemini-decision-1) scores candidate options in a single forward pass and returns probabilities and confidence scores (0 to 1) instead of generating text.
  • Google quotes 150-250 ms end-to-end latency (50-100 ms median model compute) for contexts up to 16K tokens, at $0.10 per 1M input tokens with zero output tokens billed.
  • One request takes a multimodal input of up to 256K tokens and up to 1,000 independent questions, in three types: noul (yes/no), choice (2 to 255 options) and score (ordered rubric).
  • Google says the model is nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API, and leads its benchmarks with the lowest calibration error; no benchmark values are published.
  • Gemini Notebook, Stitch, CC from Google Labs and Google Search already use the API; Gemini Notebook reports up to 40% lower latency for everyday queries.

Why it matters

Many LLM calls in real products are not writing anything. They pick a tool, route a query, check a policy or grade context, and a full generative call is slower and costlier than the job needs. Google's answer is a dedicated endpoint where the model scores a fixed set of options in one forward pass, so there is no token-by-token decoding and no output billing. The calibrated confidence score on every answer is the other selling point: code can act on clear cases and hand ambiguous ones to a fallback or a human reviewer. Google frames it as a control plane that teams build by defining options rather than training custom classifiers.

Who it affects

Developers building agents and LLM pipelines who today spend generative calls on routing, tool selection, moderation-style checks or context grading. Teams working with mixed media are also in scope, since one input can combine text, images, audio, video and PDFs, for example a call recording with account metadata or a scanned claim with damage photos. Inside Google, Gemini Notebook, Stitch, CC from Google Labs and Google Search already depend on it. Users of OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode can expect integrations as they roll out.

How to use it

Get an API key in Google AI Studio and call client.decisions.create from the google-genai SDK, with model set to gemini-decision-1 and the v1beta API version as in Google's example. Pass a shared input in the same Content block format as the Interactions API, then define a questions object. Use noul for yes/no conditions, choice for 2 to 255 labelled options with criteria, and score for an ordered rubric. Read the returned probabilities and the 0 to 1 confidence, and route low-confidence cases to a fallback or a person. Pricing is $0.10 per 1M input tokens and $0 for output; the shared input is billed once per request and each question adds only its own instruction and criteria tokens. An API no-cost tier in Google AI Studio is available for prototyping. Google points to the Doorstep Parcels, Speed Sketch and Find the Zoo Keeper demo apps and to Python, JavaScript and REST guides. Support in Gemini in Google Cloud is coming soon.

How solid is it

This is a first-party Google announcement on blog.google, so the capabilities, price and latency are Google's own statements. The price, input limits, question types and example code are concrete and checkable. The performance claims are not: Google says the model leads across its benchmark suite with the lowest calibration error, but the post gives no benchmark names, scores or calibration values, and the only named comparison is Gemini 3.5 Flash-Lite in the cost claim. The post does not explain how the nearly 5x cheaper figure was computed beyond 'across the shared benchmark suite'. The production examples are all Google products, and no external customers are named.

Risks and caveats

The stated latency figures (50-100 ms compute, 150-250 ms end-to-end) are tied to contexts up to 16K tokens, while the maximum input is 256K tokens; the post does not give latency for larger inputs. The up to 40% latency gain is for Gemini Notebook only, and its baseline is not stated. No limits or quotas of the no-cost tier are stated. No model size, architecture details or training data are disclosed. No general availability or preview status is stated, and the only date for Google Cloud support is 'coming soon'. The model only picks among options you define, so it suits bounded decisions and not open-ended generation. Google's own example runs against the v1beta API version.

“Rather than generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass, returning probabilities and confidence scores”

— Google, Gemini Decisions API announcement