Google's Gemini Decisions API scores 1,000 questions in one pass
A new endpoint built on the Gemini Decision 1 model returns probabilities and confidence scores for bounded choices instead of writing answer text.
What happened
On Nov 11, 2026, Google introduced the Gemini Decisions API, a new endpoint for closed-set decisions, meaning bounded choices with a fixed list of answers. It runs on a new model, Gemini Decision 1 (gemini-decision-1), available through the Gemini API in Google AI Studio. The launch post is credited to Min Choi, Product Manager, and Felix Yu, Principal Scientist, both at Google DeepMind.
The scope is narrow on purpose. Instead of generating answer text token by token, the model scores candidate options you supply and returns probabilities and confidence scores. Google positions it for tool selection, policy checks, model routing and context grading.
What the source confirms
- A single request carries one shared multimodal input (text, images, audio, video and PDFs, up to 256K tokens) plus up to 1,000 independent questions.
- End-to-end latency is 150-250 ms, a median of 50-100 ms of model compute, measured across multimodal inputs and contexts up to 16K tokens. Latency is said to stay nearly flat from 1 to 100 questions per request.
- Pricing is $0.10 per 1M input tokens and $0 for output tokens, with a no-cost tier in Google AI Studio for prototyping. The shared input is billed once per request, and each question adds only its own instruction and criteria tokens.
- Google says the setup is nearly 5x cheaper across a shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
- Across benchmarks for agent routing, reasoning, ranking and multimodal understanding, Google says Gemini Decision 1 leads with the lowest calibration error. No individual scores are printed in the post.
Three question types, one pass
Every question takes one of three formats. A Yes/No check estimates the probability that a condition holds, with optional true and false description criteria. A multiple choice question picks from 2 to 255 options and returns the winner plus normalised probabilities for every candidate. An ordered rubric grades an input starting at 0 and returns a continuous weighted score with per-level probabilities.
Each answer carries a confidence score from 0 to 1, which Google says lets your code route ambiguous cases to a fallback or a human reviewer. Input uses the same Content block format as the Interactions API, and resolution on image and video inputs can be tuned to control visual detail and token cost.
How Google teams are using it
Gemini Notebook uses the API as a real-time router, classifying intent in milliseconds to send everyday queries to a cheaper tier, which Google says cuts latency by up to 40% for those queries while reserving deeper reasoning models for complex tasks. The rapid canvas experiment in Stitch generates playable wireframes in seconds and applies voice or tap edits almost instantly. Google Search uses it as a safety layer for MCP tools, detecting adversarial behaviour and content within the latency envelope.
The example in the source
A code sample sends one photo of a breakroom counter with a text note about a 9 AM client meeting. Three questions run against that shared input: whether there is an uncleaned liquid spill, which cleanup action fits (wipe with cloth, dispatch janitorial, ready for meeting), and a three-level severity rubric. The response returns a probability, a chosen action and a weighted score.
Limits and open questions
Google claims leadership and the lowest calibration error but prints no per-benchmark numbers, so the margin's size is unknown. The input cap is 256K tokens, while the flat-latency claim is stated for contexts up to 16K tokens. Ecosystem integrations are described as rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with support coming soon to Gemini in Google Cloud, and no dates are given. Rate limits, regional availability and guidance on picking confidence thresholds are not covered. The source also labels the Yes/No type "noul", which may be a field name or a typo.
What it means for building apps with AI
Most AI features need many small judgments, not long prose: is this a support request or a billing question, does this image break a rule, which tool should handle this. A separate scoring endpoint with published per-request pricing suggests splitting those cheap calls away from expensive generation. Whether the probabilities hold up on your own data is something you would have to measure.
Draftship builds apps from a plain-language description and hosts them, with a built-in database and user sign-in. Paid plans add a choice of AI model, source code access, custom domains and GitHub push. It is not affiliated with Google. If you are adding model calls to an app, see adding AI to your app.
Primary sources
- Gemini Decisions API: Fast, calibrated multimodal decisions · Google AI Blog, Nov 11, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.