Multimodal · NOVEMBER 12, 2026
Gemini Decision 1 batches 1,000 checks—but intake routing needs a pilot
Google’s new API scores multimodal inputs without generating prose. Its shared-input pricing could suit customer intake, but latency claims cover smaller workloads than the advertised limits.
Google launched the Gemini Decisions API on November 11, using Gemini Decision 1 to evaluate up to 1,000 independent multimodal questions and return probabilities, not prose.
For small-business intake, that could mean checking request type, urgency and missing information together. But Google’s speed claims cover smaller workloads, and there’s no evidence here of increased leads or conversions. The earlier batched-checks breakdown explains the approach.
Pricing and input limits
Google charges $0.10 per million input tokens, without output-token charges. Shared evidence is billed once per request; questions add instruction and criteria tokens. Separate requests rebill evidence. Google lists no separate audio-transcription, OCR or thinking-token charges.
VentureBeat’s example pairs a 4,000-token record with 20 checks of 50 tokens: $0.50 per 1,000 records batched, versus $8.10 across 20 separate calls.
Checks support yes/no, multiple choice with 2–255 options, and rubrics starting at zero. Multiple choice returns every candidate’s probability; the selected probability is confidence, from 0 to 1. Rubrics return weighted scores and per-level probabilities.
Google advertises 256K-token inputs and adjustable image/video resolution; VentureBeat specifies 262,144 tokens including evidence, instructions and choices. Questions are independent. Dependent decisions need application logic or another call; replies and arbitrary invoice-number extraction need separate steps. SiliconANGLE identifies this narrower scope as the trade-off.
Availability isn’t a latency guarantee
Access uses gemini-decision-1, Google AI Studio, Python, JavaScript and REST guides with v1beta. Google offers no-cost prototyping; VentureBeat reports limits of 200 requests and 50,000 input tokens per minute. OpenRouter, LangChain and Vercel integrations are rolling out; Google Cloud support remains forthcoming.
Google DeepMind product manager Min Choi and principal scientist Felix Yu report 150–250 milliseconds end-to-end, including 50–100 milliseconds median model compute, for contexts up to 16K tokens. Near-flat latency covers 1–100 questions, not 1,000. Google’s up-to-40% reduction concerns Gemini Notebook’s internal workflow, not intake.
Benchmarks argue for supervised testing
Google’s nearly fivefold cost advantage over Gemini 3.5 Flash-Lite concerns its benchmark workload. VentureBeat lists Decision 1 at $0.88, Flash-Lite at $4.14 and Jev at $0.68. Google reports calibration error of 0.0250 versus Jev’s 0.0494. Its MMAU audio scores were 71.8% versus Flash-Lite’s 67.7%; Video-MME scores were 69.1% versus 63.3%.
VentureBeat reports Artificial Analysis Decision Index v2.1 ranked Gemini ninth of 18: 65.98 against Jev 1.14’s 66.23. Gemini recorded 211 milliseconds per task and $345 per million tasks; Jev recorded 92 milliseconds and $156. Latency combines per-test medians; costs reflect the suite’s task mix, not intake.
On AA-Paperclip’s 15 questions per document, Gemini answered everything correctly on 33% of documents, versus Jev’s 35.5% and Drex Max v1.1’s 63%. InfoWorld cautions that confidence alone doesn’t establish trustworthiness.
The practical implication: use confidence to decide what needs review, not as permission to skip validation.
A supervised pilot alongside existing follow-up could test forms, calls, images or PDFs. VentureBeat recommends tracking errors, review volume and 95th-percentile full-request latency, including media uploads.