Google Launches Gemini Decisions: Low-Cost Multimodal Scoring

Google has launched Gemini Decisions, a decision-making API powered by the Gemini Decision 1 model. The offering provides multimodal evaluation in a single pass, confidence scores for each response, and pricing at $0.10 per million input tokens, with no charges for output. According to the company, the median latency drops to 50–100 ms on the model side, and calibration error is minimal on its benchmarks.

Simple Pricing and Deployment Integrations

Gemini Decision 1 is priced at $0.10 for 1 million input tokens and $0 for output tokens, with a no-cost tier for prototyping in Google AI Studio. Shared input is only charged once per request, regardless of the number of questions asked. Integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode are being deployed, and support for Gemini in Google Cloud has been announced. Public demos are available, including Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. You can get started today with an API key obtained from Google AI Studio, with guides available for Python, JavaScript, and REST.

Single-Pass Evaluation on Text, Image, Audio, Video, and PDF

Built on Gemini's proprietary multimodal architecture, the API supports text, images, audio, video, and PDF files in a single operation. It is possible to combine multiple modalities in a single submission, such as an audio call file accompanied by account metadata, or a scanned claim in PDF format with photographs of damages, while adjusting the resolution to control both visual accuracy and token consumption. A single request allows for the analysis of varied questions and assigns each a confidence index between 0 and 1, which helps guide uncertain cases. Three types are offered: noul to estimate a binary probability with optional criteria, choix to select between 2 and 255 options with normalized probabilities, and score to apply an ordered scale starting from 0 with a weighted continuous score. The announced capacity goes up to 1,000 questions in one pass.

Announced Calibration and 50–100 ms Latency on the Model Side

According to Google, Gemini Decision 1 exhibits the lowest calibration error on dedicated test sets for agent routing, reasoning, ranking, and multimodal understanding. The analysis is performed without going through token-by-token decoding, with a median latency of 50 to 100 ms for model execution and 150 to 250 ms in total. Furthermore, the company states that the Decisions API, which does not charge for generated output tokens, costs nearly five times less than Gemini 3.5 Flash-Lite via the Interactions API on shared benchmarks.

Internal Use Cases: Routing, Design, Emojis, and Security

In Gemini Notebook, the API acts as a real-time router to direct requests to the appropriate model. The Stitch tool uses it to produce interactive prototypes of a product journey in seconds. In CC, the API instantly selects an emoji reaction from over 200 possibilities. In Search, it serves as a security layer for MCP tools, detecting adversarial behaviors and content.

Probabilistic Decisions and Usage Economy

The API aims for rapid and probabilistic decision-making in multimodal contexts, providing confidence scores per question in a single pass. It is presented as a cost-effective way to handle queries, evaluate content, or verify policies. Teams can get started immediately via Google AI Studio and rely on Python, JavaScript, and REST documentation for implementation.