Google Launches Gemini Decision 1: A Model That Answers Instead of Writing

How it unfolded

Google DeepMind product manager Min Choi and principal scientist Felix Yu announced the Gemini Decisions API on Nov 11 in the company's official blog, powered by a new model, Gemini Decision 1 (gemini-decision-1), available now through the Gemini API in Google AI Studio, with support in Gemini on Google Cloud listed as coming soon.

Instead of generating answers token by token, the model scores every candidate option in a single forward pass and returns probabilities with confidence scores from 0 to 1. Three question types are offered: yes/no, multiple choice (2 to 255 options) and scoring (an ordered rubric starting at 0, up to 255 levels). One shared input caps at 262,144 tokens and can mix text, images, audio, video and PDFs; each question sees only the shared input, not the other questions, so dependent judgments must be split across two requests by application logic.

Google says median model compute takes 50 to 100 milliseconds and end-to-end latency runs 150 to 250 milliseconds, staying nearly flat for contexts up to 16K tokens and from 1 to 100 questions per request. Pricing is $0.10 per million input tokens with no charge for output tokens, and the shared input is billed once per request; Google says it is nearly five times cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. The AI Studio no-cost tier caps at 200 requests a minute and 10,000 a day.

Google says it already uses the endpoint internally: Gemini Notebook routes queries to different model tiers and thinking levels in real time, cutting latency for everyday queries by up to 40%; a Stitch rapid-canvas experiment uses it to generate playable product-flow wireframes in seconds; CC from Google Labs uses it to pick a reaction from more than 200 emoji; and Search uses it as a safety layer for MCP tools, detecting adversarial behavior and content. Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.

What the parties said

In the blog post, Google positions the endpoint as purpose-built primitives for bounded choices, aimed at tool selection, policy checks, model routing and context grading, saying teams can ship features simply by defining options, rather than training custom classifiers or waiting on a generative call and then parsing text back into values.

On performance and calibration, Google says Decision 1 leads across its benchmark suite spanning agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. Those figures come from Google's own tests and had not been independently reproduced at the time of publication; the decision indexes maintained by Artificial Analysis and Hugging Face will be the first third-party comparisons.

The launch materials also state the limits: initial access runs through the Gemini Developer API v1beta only, and the SDK quickstarts require google-genai 3.2 or later; question instructions are text-only, and conversation history and tool-call blocks are unsupported; context caching, the Batch API, and Flex or Priority inference are unavailable. The documentation recommends setting confidence thresholds with labeled examples from a developer's own traffic and retaining a human or second check for high-stakes actions such as payouts.

The competitive picture

The decision-model category only appeared in September: TypeSafe launched Jev first on Sept 15, in early access, priced at $0.042 per million input tokens with output free; OpenAI previewed its Decisions API at DevDay on Sept 29, built on GPT-6 Luna, entering public beta on Oct 6 and reaching general availability on Oct 22; Microsoft launched Microsoft-Decision-1 on Oct 9, post-trained from Alibaba's Qwen3.5-9B and priced at the same $0.042.

Within a month, AWS, Cloudflare, Perplexity, Snowflake, Liquid AI and more than a dozen other vendors and open-source projects followed with products of the same kind, Hugging Face hosts a Decision Index, and TypeSafe calls the category System One models. Google is the fourth large vendor to ship a decisions endpoint in two months, and the latest.

On price, Google's $0.10 matches OpenAI and sits above the $0.042 charged by Jev and Microsoft. Google's difference is multimodal input: one shared input can mix text, images, audio, video and PDFs, while OpenAI's endpoint accepts text and images and Microsoft's accepts text and JSON only.

Calibration differs too: TypeSafe says Jev calibrates confidence against real outcomes with RLCD; Google says Decision 1 has the lowest calibration error, but that claim comes from Google's own benchmarks and has not been verified by a third party. OpenAI's endpoint likewise has a general-purpose model return probabilities in a single call.

Key facts

1. Model: Gemini Decision 1, model ID gemini-decision-1; it generates no free-form text and answers only bounded questions
2. Volume: Up to 1,000 independent questions per request, in three types: true/false, multiple choice with 2 to 255 options, and scoring with up to 255 levels
3. Latency: Google reports 50 to 100 milliseconds of median model compute and 150 to 250 milliseconds end to end; 1 and 100 questions take nearly the same time
4. Pricing: $0.10 per million input tokens, $0 for output; the AI Studio no-cost tier caps at 200 requests a minute and 10,000 a day; Google says it is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API
5. Context: Shared input plus all questions caps at 262,144 tokens, covering text, images, audio, video and PDFs
6. Internal use: Gemini Notebook routing (up to 40% lower latency on everyday queries), Stitch rapid canvas, CC emoji selection, Search MCP tool safety layer
7. Access: Available now through the Gemini API and Google AI Studio (v1beta); integrations with OpenRouter, LangChain, Vercel, LiteLLM and others are rolling out; Gemini on Google Cloud support is coming soon
8. Category timeline: TypeSafe Jev on Sept 15; OpenAI Decisions API previewed Sept 29, generally available Oct 22; Microsoft-Decision-1 on Oct 9; Google on Nov 11