Google Launches Gemini Decisions API for 150ms Multimodal Routing
Google's Gemini Decisions API and gemini-decision-1 model process 1,000 classification questions in 150 to 250ms at $0.10 per million input tokens.

By Clara Tan, Artificial Intelligence
4 min read
Purpose-built silicon and API for closed-set agent routing
Google DeepMind has released the Gemini Decisions API, a dedicated endpoint powered by the new gemini-decision-1 model in Google AI Studio and the Gemini API. Announced on 11 November 2026 via the Google blog, the service targets high-volume backend tasks such as support triage, content moderation, policy verification, and agentic workflow routing where teams previously relied on general-purpose chat models.
Traditional large language models generate prose or JSON one token at a time, adding decoding latency and forcing developers to write defensive parsers for malformed responses. gemini-decision-1 skips autoregressive text generation entirely, emitting zero output tokens and returning structured selections in 150 to 250 milliseconds end to end.
Specifications, question types, and token pricing
Developers can send a single multimodal payload covering text, images, audio, video, or PDFs up to roughly 262,000 tokens alongside up to 1,000 independent decision questions in one API call. According to Google's developer documentation, the endpoint supports three structured decision modes:
Binary checks (noul) for yes-or-no policy gates and binary classification.
Multiple choice (choice) supporting between 2 and 255 discrete categories for routing and tagging.
Ordered rubric scoring (score) for ranking, quality grading, and risk bands.
Google priced the API at $0.10 per 1 million input tokens, with output tokens free because none are generated. For engineering teams running multi-agent pipelines that inspect long transcripts or compliance documents before invoking heavier reasoning models, batching 1,000 checks against a shared 262K context window cuts both prompt-caching complexity and tail latency.
Where Decisions fits against Frontier and Flash models
The release sits alongside Google's broader Gemini 4 family, including the recently launched Gemini 4 Hydrogen model aimed at Search and consumer agentic workflows. While Hydrogen and Carbon handle open-ended synthesis, code generation, and tool calling, gemini-decision-1 acts as a fast deterministic gatekeeper upstream.
In production agent architectures, teams frequently burn 40% to 60% of their inference budget on repetitive guardrail checks, intent classification, and document filtering. Moving those fixed-schema checks onto a zero-output classifier reduces token bills and removes JSON parsing failures from critical production paths.
Frequently Asked Questions
How much does the Gemini Decisions API cost?
Google prices gemini-decision-1 at $0.10 per 1 million input tokens. Because the model does not generate prose, there are zero output token charges.
How many questions can one Gemini Decisions API call evaluate?
A single request can pair one multimodal input of up to roughly 262,000 tokens with up to 1,000 independent binary, multiple-choice, or rubric-scoring questions.
When should teams use gemini-decision-1 instead of Gemini 4 Flash or Hydrogen?
Use gemini-decision-1 for closed-set triage, moderation, routing, and compliance scoring where the possible answers are known in advance and latency under 250ms matters.