↓ Skip to main content
9 – 17 UHR +49 8031 3508270 LUITPOLDSTR. 9, 83022 ROSENHEIM
DE / EN
DECISION Google USA

Gemini Decision 1

Gemini Decision 1: decision model behind the Gemini Decisions API with calibrated probabilities, multimodal, 256K input, USD 0.10 per 1M input tokens.

License Proprietary
Context 262,144 Tokens
Modality Text, Image, Audio, Video, PDF → Calibrated probabilities and confidence per question

Versions

Overview of available model variants

ModelReleaseEUStrengthsWeaknessesStatus
Gemini Decision 1
11 November 2026
Multimodal: text, images, audio, video and PDFs in one shared input, resolution adjustable for images and video Up to 1,000 independent questions per request over up to 262,144 tokens of shared context Question types noul (yes/no), choice (2 to 255 options) and score (2 to 255 rubric levels), every answer with a confidence from 0 to 1 Per Google 150 to 250 ms end-to-end, nearly flat between 1 and 100 questions per request USD 0.10 per 1M input tokens, no output tokens; the shared input is billed once per request Per Google leading on its own benchmark suite with the lowest calibration error, and nearly 5 times cheaper than Gemini 3.5 Flash-Lite with the Interactions API Free tier and Decisions Playground in Google AI Studio for quick prototypes
Only through the Gemini API (v1beta) and AI Studio; Google Cloud per Google only coming soon No EU data residency: per the terms of service, data may be stored transiently in any country with Google facilities No context caching, no Batch API, no Flex or Priority inference; repeated input is billed in full every time No rationales, no text output, no tool calls; question instructions text only Proprietary, no open weights; self-hosting not possible Google publishes no individual benchmark scores; independent evaluations are not yet available Free tier with 50,000 tokens per minute, well below the 262,144-token context window
Current

Use Cases

Typical applications for this model

Model routing and choice of thinking level
Tool selection and agent control
Policy and safety checks across many rules at once
Scoring and ranking of candidates against a rubric
Classification of images, documents, audio and video
Real-time reactions to UI state and media streams

Technical Details

API, features and capabilities

API & Availability
Availability Gemini API (POST https://generativelanguage.googleapis.com/v1beta/decisions) and Google AI Studio; google-genai SDK version 3.2 or later (Python client.decisions.create, JavaScript ai.decisions.create)
Requests/Min 200 (Free), 1,500 (Tier 1), 5,000 (Tier 2), 15,000 (Tier 3)
Tokens/Min 50,000 (Free), 1M (Tier 1), 4M (Tier 2), 10M (Tier 3)
Latency (TTFT) Per Google 150 to 250 ms end-to-end, median 50 to 100 ms model compute
Features & Capabilities
Vision File Upload
Training & Knowledge
Knowledge Cutoff not documented
Fine-Tuning Not available
Language Support
Best Quality English
Supported No language list published
Google does not document language coverage; for German-language decision tasks, run your own evaluation before deployment

Hosting & Compliance

GDPR-compliant hosting options and licensing

License & Hosting
License Proprietary (Google, API only)
Security Filters Input safety filters; blocked requests return HTTP 400 with an error code (safety, prohibited_content, blocklist, image_safety, content_blocked)
Cloud Only

Gemini Decision 1 is a decision model from Google, available since 11 November 2026 through the new Gemini Decisions API. It belongs to the category of System One models that we explained in our post Jev by TypeSafe: the AI model that doesn’t write text: instead of formulating an answer token by token, the model scores predefined answer options in a single pass and returns a calibrated probability for each. Google positions the API for closed-set decisions such as tool selection, policy checks, model routing and context grading.

Gemini Decision 1 processes text, images, audio, video and PDFs in one shared input and answers up to 1,000 questions per request. This page is based on the announcement on the Google blog by Min Choi (Product Manager) and Felix Yu (Principal Scientist) of Google DeepMind, the Decisions API documentation, the model page, the pricing page, the rate limits and the Gemini API terms of service. This overview reflects the state as of 11 November 2026.

What is Gemini Decision 1?

Per Google, Gemini Decision 1 (model ID gemini-decision-1) is built on Gemini’s natively multimodal architecture. Google does not name a base model, a parameter count or a knowledge cutoff. The model generates no text tokens: the output consists solely of probabilities and confidence values, and the output token count is always zero.

Each call sends a shared input and a list of questions. Per the documentation, every question is evaluated independently against the shared input: questions cannot see each other, and adding or removing a question does not change the other answers. The application then combines the answers in its own code, for example with thresholds and if conditions.

Google does not state an explicit release stage (preview or GA) in the documentation or on the pricing page. The changelog lists the launch on 11 November 2026 as “Released”. The model ID follows the naming scheme of stable Gemini models without a preview suffix, but the endpoint sits in API version v1beta.

Question types and answers

As with Jev and Microsoft-Decision-1.1, the question types are called noul, choice and score:

Question typePurposeReturns
noul (yes/no)Does a statement hold?probability of yes (0.01 to 0.99), confidence
choice (selection)Pick one option from a fixed listselected option, probability per option (sums to 1), confidence
score (rubric)Rate the input on an ordered scaleprobability-weighted value, probability per level, legend, confidence

A choice question allows 2 to 255 options, a score question 2 to 255 levels starting at 0; Google recommends 2 to 10 levels for most rubrics. The confidence value from 0 to 1 indicates how likely the top answer is to be right. Applications can use it to decide which cases proceed automatically and which go to a human or a larger model. Google explicitly advises setting thresholds with labeled examples from your own traffic. The API sorts option keys alphabetically before evaluation; for a catch-all option Google therefore recommends a key such as “other”.

Calls go to POST https://generativelanguage.googleapis.com/v1beta/decisions or through the google-genai SDK version 3.2 or later (client.decisions.create(...) in Python, ai.decisions.create(...) in JavaScript). The input field uses the same Content block format as the Interactions API.

Inputs, limits and restrictions

LimitValue per the documentation
Context window262,144 tokens for shared input and all questions
Questions per request1 to 1,000
Options per choice question2 to 255
Levels per score question2 to 255 (2 to 10 recommended)
Inline dataup to 100 MB per request, PDFs up to 50 MB
External URLsup to 15 MB per URL
Input formatstext, images (including PNG, JPEG, WebP, HEIC), audio (including WAV, MP3, FLAC, Opus), video (including MP4, MOV, WebM), PDF

For images and video, you can choose the resolution (low, medium, high, and for images also ultra_high) to control visual detail and token cost. The documentation lists several restrictions:

  • Question instructions accept text only; media belongs in the shared input.
  • YouTube URLs and direct Cloud Storage URIs (gs://) are not supported; media comes in via the Files API, public or signed HTTPS URLs, or inline as base64.
  • Conversation history, thought and tool-call blocks are not accepted as input; serialize multi-turn context as text.
  • There is no context caching, no Batch API and no Flex or Priority inference. Repeated input counts in full on every request.
  • If the safety filters block an input, the API returns HTTP 400 with an error code instead of an answer.

The model produces no rationales. If you need a traceable explanation, combine the Decisions API with a language model or document the decision logic in the workflow itself.

Benchmarks and examples per Google

Per Google, Gemini Decision 1 leads a benchmark suite covering agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. Google does not publish individual scores or the list of comparison models in the blog post. Across the same benchmark suite, Gemini Decision 1 with the Decisions API is, per Google, nearly 5 times cheaper than Gemini 3.5 Flash-Lite with the Interactions API.

On latency, Google states 150 to 250 ms end-to-end and a median of 50 to 100 ms of pure model compute. Per Google, these values apply to multimodal inputs and contexts up to 16K tokens and stay nearly flat between 1 and 100 questions per request.

From internal use at Google, the blog post gives four examples:

  • Gemini Notebook routes incoming queries via the Decisions API to the right model tier and thinking level, per Google with up to 40 percent lower latency for everyday queries.
  • Stitch generates playable wireframes in its “rapid canvas” experiment and applies edits made by voice or tap.
  • CC from Google Labs picks the right emoji reaction from more than 200 options.
  • Google Search uses the API as a safety layer for MCP tools to detect adversarial behavior and content.

All figures come from Google; independent evaluations are not yet available.

Pricing and rate limits

ModelInput (USD per 1M tokens)OutputFree tier
Gemini Decision 10.10 (text, image, video, audio, PDF)no output tokensfree of charge in Google AI Studio

Prices per Google, as of November 2026. The shared input is billed once per request; each question only adds its own instruction and criteria tokens. Per Google, there are no charges for thinking tokens, separate audio transcription or OCR. Because context caching is not available, applications with large, recurring inputs should bundle as many questions as possible into one request.

TierRequests/minuteTokens/minuteRequests/day
Free20050,00010,000
Tier 11,5001,000,00050,000
Tier 25,0004,000,000unlimited
Tier 315,00010,000,000unlimited

Google notes that rate limits are not guaranteed and actual capacity may vary.

EU availability and GDPR

Gemini Decision 1 is available exclusively through the Gemini API and Google AI Studio. Google announces support for Gemini in Google Cloud, that is the Gemini Enterprise Agent Platform with its regions and residency options, as coming soon, without a date. Per Google, integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode are rolling out; the announced OpenRouter model page was not yet reachable on 11 November 2026.

The Gemini API terms of service apply:

Free tier (Unpaid Services)Paid tier (Paid Services)
Used to improve productsyes, including possible human reviewno
Processingunder Google’s privacy policyunder the Data Processing Addendum (Google as processor)
Loggingno retention period statedfor a limited period, solely for abuse detection and legally required disclosures
Storage locationno EU commitmenttransient storage or caching possible in any country in which Google or its agents maintain facilities

Two rules matter for companies in Europe. First, per the terms, the paid-services data rules also apply to Google AI Studio and the unpaid quota of the Gemini API for users in the EEA, Switzerland and the United Kingdom. Second, applications made available to users in the EEA, Switzerland or the United Kingdom may only use Paid Services.

There is no EU data residency. Google does not document whether the Decisions API stores requests by default like the Interactions API; the request reference lists no store parameter. For guaranteed zero data retention or enterprise data processing agreements, Google’s ZDR page points to the Agent Platform, where Gemini Decision 1 is not yet available. There are no open weights, so running it on your own infrastructure is not possible. More on the difference between the Gemini API and the Agent Platform is on the Google Gemini page.

For workflows with personal data, we currently recommend alternatives with an EU option: Microsoft-Decision-1.1 as Data Zone Standard in an EU region, or self-hosted open models such as Cloudflare Clef, Julia-1 or GLiNER2.5-Decide.

Where it fits among decision models

With Gemini Decision 1, Google now follows OpenAI and Microsoft with a decision API of its own. Compared with the other models in this category:

  • Jev from TypeSafe shaped the category and the question types noul, choice and score, and runs as a hosted API from the US.
  • Clef from Cloudflare is open (Apache 2.0), multimodal and can be self-hosted.
  • Laya, Julia-1 and GLiNER2.5-Decide cover compact, CPU-capable variants for self-hosting.
  • Microsoft-Decision-1.1 processes text, JSON and images with 32K tokens of context, costs USD 0.042 per 1M input tokens (USD 0.05 in the EU Data Zone) and runs in Microsoft Foundry with an EU data zone option.
  • OpenAI offers a Decisions API under the same name, based on GPT-6 Luna with up to 400 questions per request (details on the OpenAI GPT page).
  • Gemini Decision 1 combines text, image, audio, video and PDF in one input with 262,144 tokens of context and up to 1,000 questions per request, but is only usable through the Gemini API without EU data residency.

Our recommendation

Gemini Decision 1 is relevant when decisions depend on images, documents, audio or video: for example checking damage photos together with a PDF claim, classifying call recordings, or running many policy checks on the same input in one request. The free tier and the Decisions Playground in Google AI Studio work well for prototypes, as long as no personal or confidential data is involved.

For production use with personal data, we advise waiting for the announced availability in Google Cloud and then reviewing its regions and residency options. Until then, Microsoft-Decision-1.1 with an EU Data Zone or self-hosted open decision models are the obvious choice. Regardless of the provider, we recommend calibrating thresholds with your own test cases, including German-language ones, and treating inputs as untrusted: Google itself notes that user text can argue for its own classification and recommends a second check for high-stakes actions. In CompanyGPT and the AI Gateway, decision models can serve as an upstream routing and checking stage; we support selection, evaluation and integration.

Related decision models

The decision model (System One) category also includes Microsoft-Decision-1.1, Cloudflare Clef, Laya, Jev, Julia-1 and GLiNER2.5-Decide. OpenAI is moving in the same direction with the Decisions API based on GPT-6 Luna; details are on the OpenAI GPT page. Google’s generative Gemini models are listed under Google Gemini.

Cost estimation for this model

For up-to-date token pricing, model variants and EU availability, see our sister project ai-prices.eu. It helps you compare and estimate the operational cost of leading AI models for your specific use case.

Compare prices on ai-prices.eu

ai-prices.eu is a project by innFactory AI Consulting GmbH and provides transparent cost estimates for leading AI models.

Frequently Asked Questions

What is Gemini Decision 1?

Gemini Decision 1 (model ID gemini-decision-1) is a decision model from Google, released on 11 November 2026 together with the new Gemini Decisions API. Instead of generating text, it scores predefined answer options in a single forward pass and returns calibrated probabilities plus a confidence value from 0 to 1. It processes text, images, audio, video and PDFs and is available through the Gemini API and Google AI Studio.

What is the Gemini Decisions API used for?

Google names tool selection, policy checks, model routing and context grading as typical tasks. Per the documentation, it also covers filtering and ranking candidates against a rubric, agent decisions (which tools to use, whether to compact, escalate or stop) and reacting to UI state, audio or video frames in real time. For generated text, extraction to your own JSON schema, tool calls and multi-step reasoning, Google points to the Interactions API.

How much does Gemini Decision 1 cost?

Per Google, the model costs USD 0.10 per 1 million input tokens on the paid tier, and no output tokens are produced. The shared input is billed once per request; each question only adds its own instruction and criteria tokens. On the free tier in Google AI Studio, use for prototyping is free of charge. There is no context caching and no Batch API, so repeated input is billed in full on every request.

How fast is Gemini Decision 1?

Google states 150 to 250 ms end-to-end latency and a median of 50 to 100 ms of pure model compute, measured with multimodal inputs and contexts up to 16K tokens. Per Google, latency stays nearly flat between 1 and 100 questions per request. Gemini Notebook uses the API as a router and, per Google, achieves up to 40 percent lower latency for everyday queries.

What are the limits of the Gemini Decisions API?

Per the documentation, the shared input and all questions share a context window of 262,144 tokens. A request contains 1 to 1,000 questions, a choice question 2 to 255 options and a score question 2 to 255 levels (2 to 10 recommended). Rate limits range from 200 requests and 50,000 tokens per minute on the free tier to 15,000 requests and 10 million tokens per minute on Tier 3; Google does not guarantee these values.

Can Gemini Decision 1 be used GDPR-compliantly?

Only with restrictions. The model runs exclusively through the Gemini API, for which Google offers no EU data residency: per the terms of service, prompts and responses may be stored transiently in any country in which Google or its agents maintain facilities. On the paid tier, Google does not use the data to improve its products and processes it under the Data Processing Addendum. On the free tier, data is used to improve products, but per the terms the paid-tier data rules also apply there for users in the EEA, Switzerland and the United Kingdom. Google announces support in Google Cloud as coming soon.

How does Gemini Decision 1 differ from Microsoft-Decision-1.1 and the OpenAI Decisions API?

All three are proprietary decision models that do not bill output tokens. Gemini Decision 1 is the only one of the three that also processes audio, video and PDFs; it accepts up to 262,144 tokens and 1,000 questions per request and costs USD 0.10 per 1M input tokens, but runs only through the Gemini API without EU data residency. Microsoft-Decision-1.1 processes text, JSON and images with 32K tokens of context, costs USD 0.042 (USD 0.05 in the EU Data Zone) and can be deployed in Microsoft Foundry as an EU Data Zone. The OpenAI Decisions API uses GPT-6 Luna, answers up to 400 questions per request and also costs USD 0.10 per 1M input tokens.

Consultation for this model?

We help you select and integrate the right AI model for your use case.