Google has introduced the Gemini Decisions application programming interface (API) for fast choices with clearly defined boundaries. Although it is aimed directly at providers of digital services, it could affect how quickly and cheaply you receive automated checks, rankings, or recommendations. In parallel, Google is bringing its Artificial Intelligence (AI) to additional models and Android devices.
What is the Gemini Decisions API?
The Decisions API announced by Google is powered by a new model called Gemini Decision 1. Unlike a conventional chatbot, it does not compose an answer sentence by sentence. It scores predefined options and returns probabilities and confidence values, which indicate how certain the model is about its assessment.
The system supports three types of tasks: yes-or-no decisions, choices among several options, and ratings based on a defined scale. According to Google, one request can contain up to 1,000 independent questions. The shared source material can include text, images, audio, video, and PDF documents.
That may sound less impressive than a chatbot producing a long report, but it is more suitable for many digital workflows. A service could check whether content complies with a policy, send a request to the most appropriate model, or choose among several tools. Google specifically names policy checks, tool selection, model routing, and context grading as intended uses.
The API is available through Google AI Studio. According to The Decoder’s report, Google already uses the technology internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search. Integrations with OpenRouter, LangChain, Vercel, and other platforms are being rolled out.
How fast and reliable are its decisions?
Google claims an end-to-end latency of 150 to 250 milliseconds for a request. The model scores the supplied options in one pass instead of generating text and then having another process interpret it. That can make a noticeable difference when a service needs to conduct many small checks, because the shared input is processed only once.
Independent technologist Simon Willison provides a concrete example. He gave Gemini Decision 1 a video of a stomping robot and simultaneously asked whether it contained a dragon or a robot and whether the scene was in France or the United States. The model assigned probabilities of about 0.92 to both the robot and the United States, but avoided absolute values of zero or one. Willison’s hands-on test therefore demonstrates both video processing and the probabilistic nature of the answers.
Those scores do not guarantee a correct decision. A probability of 0.92 does not automatically mean that the result is reliable enough for every use. Safety, access, and policy checks still need thresholds and fallback procedures, such as sending low-confidence cases to a human reviewer.
MarkTechPost reports that, according to Google, Gemini Decision 1 leads seven of nine accuracy benchmarks in Google’s comparison and has the lowest calibration error. Its strongest cited result is 92.2 percent on the public portion of JevBench, while its Macro-F1 score on Amazon ESCI is a much weaker 58.7 percent. It also trails Jev 1.14 on When2Call, scoring 79.9 percent compared with 82.0 percent. These figures come from Google’s evaluation and have not been independently verified by the supplied sources.
What do the fast decisions cost?
Google charges $0.10 per million input tokens. Tokens are small units into which a model divides content for processing; other media are also counted as input for billing purposes. There are no output-token charges because Gemini Decision 1 does not generate response text.
The shared input is billed once within a request, even when several questions are applied to it. The total context limit is 256,000 tokens across the input and questions. According to Google, processing time remains nearly flat for inputs up to 16,000 tokens and between one and 100 questions.
The price is low, but it is not the lowest available. Simon Willison lists the competing Jev decision model at $0.042 per million input tokens, while the referenced OpenAI model matches Google’s $0.10 price. Jev is limited to text and JSON, a structured data format; OpenAI handles text and images, while Gemini also supports audio, video, and PDFs.
The distinction from a general-purpose language model becomes clearer in the context of Google’s other Gemini update. Gemini 4 Carbon received a wider release, according to Willison, and launched at $2 per million input tokens and $10 per million output tokens. A price increase to $4 and $20 respectively was scheduled for December 1. Carbon can generate complex material, while Decision 1 serves the much narrower and cheaper role of a selection engine.
Pros and Cons of the Gemini Decisions API
Pros:
- Speed – Google quotes 150 to 250 milliseconds for a complete request involving bounded decisions.
- Multiple media types – Text, images, audio, video, and PDFs can all provide the shared basis for a decision.
- Predictable costs – Only input is billed, and up to 1,000 questions can be included in one pass.
- Graduated answers – Probabilities and confidence values make it possible to handle uncertain cases separately.
Cons:
- Narrow tasks – The model can only choose among predefined options or rate material using an established scale.
- No guarantee of accuracy – Reported benchmark results vary considerably, and Google’s performance claims have not been independently verified.
- Cloud dependency – Gemini Decision 1 runs only through Google’s hosted API, with no open model weights or self-hosting option.
- Not the cheapest provider – According to Willison, Jev costs less when a service only needs to process text or structured data.
What does the broader Gemini rollout mean for you?
As a beginner, you are unlikely to operate the Decisions API directly. A sensible first step is to check whether services clearly disclose automated assessments and let you correct a wrong decision. In everyday use, the technology might work in the background by sending your request to the right model or checking whether uploaded content complies with a set of rules.
Advanced users can pay closer attention to the probabilities a service returns. Instead of automatically accepting every score above 0.5, uncertain cases can be directed to a human reviewer or a more capable model. This preserves the speed advantage of Decision 1 without treating statistical confidence as certainty.
The expansion is not limited to the API. Google is bringing Gemini Intelligence to additional Android devices, including the Motorola Signature 27 and several Pixel 10 variants. A feature called Rambler turns spoken thoughts into organized text in real time. Task automation for activities such as ordering food, hailing rides, and booking reservations is also expanding to the United Kingdom, India, and Japan.
The situation is less clear for Switzerland. The sources do not announce a Swiss rollout of the new Android automation, identify supported Swiss languages, or explain whether personal information is processed locally. For users in Switzerland, the broader device rollout therefore does not automatically mean that every described feature is immediately available.
Together, Decision 1, Gemini 4 Carbon, and the additional Android features show Gemini becoming more divided by purpose: an expensive and versatile model for complex generation, an inexpensive specialist for rapid decisions, and device features for routine tasks. That structure could make AI services more efficient and less conspicuous in daily use. The unresolved risk is how reliably these automated decisions perform outside Google’s tests and how openly affected services handle errors, data processing, and regional restrictions.
Sources
- Gemini Decisions API: Fast, calibrated multimodal decisions – Google Blog, 2026-11-11
- Gemini Decisions API – Simon Willison’s Weblog, 2026-11-11
- Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens – MarkTechPost, 2026-11-11
- Gemini Decisions: Google bringt spezialisiertes KI-Modell für schnelle Entscheidungen statt Textgenerierung – The Decoder, 2026-11-11
- llm-gemini 0.36 – Simon Willison’s Weblog, 2026-11-10
- Gemini Intelligence: New features, on more devices – Google Blog, 2026-11-10
Image: Mustafa ezz via Pexels