Infrastructure · NOVEMBER 12, 2026
Gemini Decision 1 scores 1,000 checks per call—but latency has limits
Google’s probability-returning API costs $0.10 per million input tokens. Its potential small-business use is qualifying and routing inbound inquiries, not writing outreach.
Google introduced Gemini Decision 1 through its new Gemini Decisions API on November 11, offering scores for up to 1,000 independent questions against shared multimodal input. Rather than generating prose, the model evaluates predefined options, at $0.10 per million input tokens with no output-token charges, according to Google.
The latency claim needs a narrower reading than the advertised question ceiling. Google reports 150–250 milliseconds end-to-end and 50–100 milliseconds for median model computation, but its performance description doesn’t establish those timings at 1,000 questions.
What the latency figures cover
Google describes performance for contexts up to 16K tokens, with latency remaining nearly flat between one and 100 questions. That’s not evidence that the maximum question load has the same response time.
The company says Gemini Decision 1 scores candidate options in a single forward pass, avoiding token-by-token decoding. Questions can mix yes/no checks, multiple choice with two to 255 candidates, and ordered rubric scores.
Shared evidence is billed once per request; each question adds its own instruction and criteria tokens. For operators comparing bills, that distinction matters alongside the headline input rate. The desk’s Gemini API pricing coverage provides related pricing context.
VentureBeat notes that questions are independent. Dependent decisions require application logic or another call, so batching checks isn’t the same as executing a branching workflow.
Where inbound lead routing fits
For a small-business owner with a developer or software provider, a hypothetical use is classifying incoming prospect messages, call recordings, images or documents, then routing promising or uncertain inquiries for human follow-up.
Google returns confidence scores from 0 to 1 and explicitly describes sending ambiguous cases to a fallback or human reviewer. That could reduce manual triage before a sales conversation. It doesn’t itself find prospects, write outreach or demonstrate conversion gains.
Dependent decisions require application logic or another call, so batching checks isn’t the same as executing a branching workflow.
Input support also depends on the integration. Vercel says AI Gateway supports PDF, audio and video through AI SDK 7, while its TypeSafe-compatible and Decisions APIs accept text and images. Developers need to match the intended evidence format to the interface they actually use.
What to measure before adoption
The 1,000-question ceiling is a capacity claim, not a lead-qualification accuracy result. VentureBeat recommends running a candidate alongside the existing workflow, using labeled examples that include rare cases and poorly captured media, then measuring review rates, errors and tail latency.
For an inbound-routing pilot, cost per correctly routed inquiry would also be a useful measure. The practical decision is whether the model routes inquiries accurately while leaving an acceptable human-review workload, not simply whether it can score many questions in one call.