- Google, new newly developed dedicated AI model “Gemini Decision 1 (gemini-decision-1)” powered new endpoint “Gemini Decisions API” announced.
- A dedicated API designed for high-speed decision-making tasks with predetermined options, such as tool selection, policy checking, model routing, and context evaluation.
- Unlike traditional generative AI models that generate answer text one token at a time, it adopts a mechanism where candidate options are scored directly in a single forward pass, returning probabilities and a confidence score from 0 to 1.
On Wednesday, November 11, 2026 local time, Google announced the “Gemini Decisions API,” a new endpoint powered by its newly developed dedicated AI model, “Gemini Decision 1 (gemini-decision-1).”
The “Gemini Decisions API” is a specialized API designed for high-speed decision tasks with pre-determined options, such as tool selection, policy checking, model routing, and context evaluation. It can simultaneously evaluate up to 1,000 independent questions against a single multimodal input combining text, images, audio, video, and PDFs of up to 256K tokens.
Unlike traditional generative AI models that generate response text one token at a time, the “Gemini Decisions API” scores candidate options directly in a single forward pass, returning probabilities and a confidence score from 0 to 1. This keeps end-to-end latency at an ultra-low 150 to 250 ms (with model computation itself having a median of 50 to 100 ms).
The “Gemini Decisions API” is already available in “Google AI Studio” as a “v1beta” endpoint of the “Gemini API,” complete with a free tier for developer prototyping. Ecosystem integrations such as “OpenRouter,” “LangChain,” “Vercel,” “LiteLLM,” “Pydantic AI,” “Pipecat,” “LiveKit,” and “OpenCode” will also be rolled out sequentially, with support on Google Cloud’s “Gemini” coming soon.
Three Decision Types and Exceptional Cost Efficiency
The “Gemini Decisions API” allows you to evaluate and mix three different decision types within a single request.
The first is “noul,” a Yes/No judgment that estimates the probability of a condition being met; the second is “choice,” a multiple-choice option returning the most likely candidate and normalized probabilities for each among 2 to 255 choices; and the third is “score,” an evaluation score that grades inputs using an ordered rubric starting from 0, returning a continuous weighted score and the probability of each level. Because every answer comes with a confidence score, it is easy to implement logic such as routing ambiguous decisions to a fallback process or a human reviewer.
Furthermore, because it is designed to generate zero output tokens, the API is priced at $0.10 per 1 million input tokens, with a $0 charge for output tokens. Shared multimodal inputs are billed only once per request, and adding questions only incurs charges for the instruction and evaluation criteria tokens of each added question.
Since there are no additional fees for audio transcription or OCR, it can be used at approximately one-fifth the cost of using “Gemini 3.5 Flash-Lite” with the “Interactions API.”
Use Cases in Google’s Internal Products
The “Gemini Decisions API” is already in practical use as a low-latency control plane across multiple products within Google.
For example, in “Gemini Notebook,” it is employed as a dynamic model router that classifies incoming query intent in milliseconds and instantly routes it to the optimal model tier and thinking level, reducing daily query latency by up to 40%. It is also utilized in the UI design tool “Stitch”‘s fast canvas experimental feature for wireframe generation within seconds and instant editing via voice or taps, as well as for instantaneously selecting over 200 emoji reactions in Google Labs’ “CC,” and even as a safety layer for adversarial content detection in Google Search’s MCP tools.
Source: Google (https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/)