2026-11-11 · ← News
Gemini Decisions API turns a generative model into a fast decision node
Google has launched Gemini Decisions API for closed-set choices over text, images, audio, video and PDFs. Up to 1,000 questions per request, 150 to 250 ms latency and a price of $0.10 per million input tokens make it a cheap control layer for agents.
The image could not be loaded.
Google has released Gemini Decisions API with the Gemini Decision 1 model. A single multimodal input can be tested with up to 1,000 independent questions across three answer types: yes or no, a choice among 2 to 255 options and an ordered rubric score. The response contains probabilities and confidence scores rather than a generated paragraph.
Google reports 150 to 250 ms end-to-end latency and a price of $0.10 per million input tokens. Output tokens are free. The API is available through the Gemini API in Google AI Studio, including a no-cost tier for prototypes.
One cheap pass can replace a chain of small classifiers
The same input can be checked against many rules in one request. An agent can select a tool, grade context and decide whether an ambiguous case should go to a person at the same time. Google already uses the API for model routing, reaction selection and safety checks on MCP tools.
Multimodality is the important part. The decision layer can inspect a call recording, account metadata and a scanned document together without a separate transcription or OCR bill.
Agents gain a middle layer between fixed rules and full LLM calls
Teams have often had to choose among a fixed rule, a custom classifier and a more expensive generative model. Gemini Decision 1 offers another option: a bounded output with probabilities over the same input types handled by a multimodal model.
That can simplify agent architecture. Instead of generating text and parsing it afterward, an application receives a decision it can compare with a threshold and audit. The value sits in thousands of small workflow choices, not in a smarter conversation.
A probability still does not guarantee a correct decision
A number between 0 and 1 can look more precise than it is. Calibration depends on the actual data, the wording of criteria and the cost of an error. Google's own benchmarks do not replace evaluation on a team's documents, languages and edge cases.
The model does not provide governance either. Teams must set human-review thresholds, log decisions and handle model changes. A cheap wrong decision is still wrong.
Thresholds, audits and production results will decide adoption
The first useful signal will be stable calibration on production data. The second will be whether teams can change models or policies without rebuilding the control layer.
If integrations in LangChain, Vercel AI SDK, LiteLLM and other tools hold up, decision endpoints may become routine agent infrastructure. Error rates and the volume of human escalations will show whether the API saves work or merely creates a faster repair queue.
Lilith's verdict
An agent making a thousand decisions a minute does not need to write an essay each time. Gemini hands it a cheap ballot, but the team still has to write the counting rules.