Google DeepMind launches Gemini Decisions API for bounded AI choices

Overview

Updated 35 minutes ago

Google's Gemini API can now answer up to 1,000 yes/no, multiple-choice and rubric questions about one shared input without writing a sentence. The Gemini Decisions API, powered by Gemini Decision 1, scores the options in a single pass and returns probabilities and confidence scores.

The endpoint is built for software control flow: selecting a tool, routing a request, checking a policy or grading context. Google prices it at $0.10 per million input tokens and bills no output tokens, making a decision call closer to a classifier than a chat completion.

Why it matters

Cheap, fast decision calls let agents check many small questions before acting, changing how AI systems are built and billed.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

1,000

Questions per request

A single request can mix yes/no, choice and rubric questions against one shared input.

262,144

Context window in tokens

Covers the shared input and all questions in a request.

150–250 ms

End-to-end latency

Google reports median latency for multimodal inputs up to 16K tokens.

$0.10

Input price per 1M tokens

The shared input is billed once per request, no matter how many questions are evaluated.

Output-token price

Google bills no output tokens for Decisions API calls.

Nearly 5x

Cheaper than Gemini 3.5 Flash-Lite

Google's benchmark comparison against Gemini 3.5 Flash-Lite with the Interactions API.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Choi

Product Manager, Google DeepMind

Announced the Gemini Decisions API

Felix Yu

Principal Scientist, Google DeepMind

Co-authored the Gemini Decision 1 launch announcement

Organizations Involved

Google DeepMind

AI company

Developer of Gemini Decision 1 and the Decisions API

Google's AI research lab built the Gemini Decision 1 model and the Decisions API endpoint.

Google

Technology company

Offers the Decisions API through the Gemini API and Google AI Studio

Google operates the Gemini API and AI Studio, where the Decisions API is available.

Timeline

1 event Latest: Today

November 2026
Google launches Gemini Decisions API
Today Product launch

Google DeepMind releases a Gemini API endpoint that scores up to 1,000 closed-set questions per request. Every answer includes a confidence score.
Today

Scenarios

Gemini Decision 1 reaches general availability

Likely Resolves by May 11, 2027

Discussed by: Unite.AI, which notes initial access is limited to the v1beta Gemini Developer API

Google launched the model in v1beta. If adoption holds, the next step is a general-availability listing in the Gemini API changelog, which would remove the beta label and widen production use.

Google adds batch and caching to the Decisions API

Possible Resolves by May 11, 2027

Discussed by: Google's API documentation, which lists both as unavailable at launch

The launch docs say context caching and Batch API are unsupported. A later update could add them, making high-volume decision workloads cheaper to run.

Agent frameworks add first-class decision calls

Possible Resolves by May 11, 2027

Discussed by: Vercel's changelog and OpenRouter's model listing, which show early gateway integrations

Those integrations route Gemini Decision 1 through existing model gateways. A stronger outcome is a framework-level primitive for yes/no, choice and score calls, which would push decision APIs into standard agent stacks.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

June 2023

OpenAI function calling (2023)

OpenAI added function calling to GPT-3.5 and GPT-4. Developers could describe tools in a request and get JSON arguments back, so a model could trigger code instead of only writing text.

Then

Apps began using models to choose tools and fill parameters.

Now

Tool calling became a standard part of LLM APIs.

Why this matters now

The Decisions API pushes the same idea further: it scores bounded choices directly, without generating a text answer to parse.

August 2024

OpenAI Structured Outputs (2024)

OpenAI introduced Structured Outputs, which made model responses follow a supplied JSON Schema. The feature reduced parsing failures for developers who needed typed data.

Then

Developers could rely on schemas for extraction and workflow steps.

Now

Schema-constrained generation became common across model providers.

Why this matters now

Structured Outputs still generate tokens. Gemini Decision 1 instead returns probabilities over fixed options in one pass.

October 2018

Google BERT (2018)

Google open-sourced BERT, a language model that could be fine-tuned for classification and question-answering tasks. It improved search understanding and became a base for many NLP systems.

Then

Developers used BERT for sentiment, tagging and search ranking.

Now

Pre-trained models became the default starting point for language classification.

Why this matters now

BERT showed that scoring labels can be cheaper than generating text. Decisions API packages that idea as a managed multimodal endpoint.