Skip to main content
Back to timeline
AISource publication:

Gemini Decisions API turns bounded choices into one forward-pass score: multimodal input, up to 1,000 questions, 150-250 ms end to end

Synopsis

Google introduced the Gemini Decisions API and the gemini-decision-1 model for closed-set decisions such as tool selection, policy checks, model routing and context grading. Instead of generating answer text token by token, it scores candidate options in a single forward pass and returns probabilities and confidence scores, with one request carrying a shared input of text, images, audio, video and PDFs (up to 256K tokens) plus up to 1,000 independent questions. Google states 150-250 ms end-to-end latency, priced at $0.10 per 1M input tokens with no output tokens billed.

Source-provided article image: Gemini Decisions API: Fast, calibrated multimodal decisions
Google AI 与 Gemini 产品博客

Interpretation

The Gemini Decisions API lets a caller submit a bounded set of options together with multimodal input and get back a probability for each candidate plus a confidence score from 0 to 1, rather than having the model write out an answer token by token. Earlier generative calls had to decode answer text, billed output tokens, and grew in latency with generation length; the new endpoint turns a decision into scoring over a candidate set. Google reports 150-250 ms end-to-end latency, with median model compute of 50-100 ms, across multimodal inputs and contexts up to 16K tokens, staying nearly flat from 1 to 100 questions per request.

A single request can evaluate several questions at once, mixing yes/no checks, multiple choice and ordered rubric scoring. The three decision types share one input, which is billed once per request, while each question adds only its own instruction and criteria tokens. The yes/no type returns the probability that a condition holds, with optional true and false description criteria; multiple choice returns the winner and normalized probabilities for every candidate among 2 to 255 choices; the score type returns a continuous weighted score and per-level probabilities on an ordered rubric starting at 0.

On benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, Gemini Decision 1 leads across the suite and records the lowest calibration error. Google also gives a cost comparison: across the shared benchmark suite, Gemini Decision 1 with the Decisions API is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API. That gap comes from zero output tokens billed and $0.10 per 1M input tokens; teams inside Google already use it as a low-latency control plane, with Gemini Notebook routing queries in real time and Google reporting up to 40% lower latency for everyday queries.

Perspective

The report covers the interface capabilities, pricing and internal uses Google published, for agent and product engineers who need low-latency bounded decisions; the stated latency figures correspond to multimodal inputs and contexts up to 16K tokens.

Google says Gemini Decision 1 leads the benchmark suite with the lowest calibration error, but the material gives no per-benchmark scores or error values; the latency figures correspond to contexts within 16K tokens, and behavior on longer contexts is not described in the material.