Google launches Gemini Decisions API and Gemini Decision 1: up to 1,000 bounded questions in one pass, $0.10 per 1M input tokens

Why it matters

【Key points】

· Google posted the Gemini Decisions API on blog.google at 2026-11-11 18:00 UTC, powered by Gemini Decision 1 (gemini-decision-1).

• · It scores a shared multimodal input (text, images, audio, video, PDFs, up to 262, 144 tokens) in one pass, answering up to 1,000 bounded questions with probabilities and confidence, and zero text tokens.
• Types: noul (yes/no, 0.01–0.99), choice (2–255 options), and score.

• · Official latency: 150–250 ms end-to-end (median 50–100 ms of model compute).
• Price: $0.10 per 1M input tokens, $0 output; shared input billed once.
• AI Studio has a no-cost tier.

• · Google’s chart (not an independent index): JevBench public split 92.2% vs TypeSafe Jev 1.14 at 90.0% and OpenAI Decisions API (GPT-6 Luna) at 90.9%.
• Top-1 ECE on 7 shared evals is 0.0250 (Jev 0.0494, Luna 0.0721, Flash-Lite 0.2104).
• Those 7 cost $0.88 vs Flash-Lite $4.14 (~4.7x).
• Jev leads on Mind2Web (86.8% vs 86.2%) and When2Call (82.0% vs 77.9%) and is cheaper ($0.68).

• · Available now via the Gemini API (v1beta, SDK 3.2+) and AI Studio.
• Cloud is “coming soon”.
• No open weights; Batch, Caching and Thinking are unsupported.

【Our read】

· Background: a specialized decision API for agent hot paths, not a new frontier LLM. OpenAI, Liquid, Perplexity and Baseten have recently shipped similar products.

• · Comparison: same class as the OpenAI Decisions API (gpt-6-luna), with audio/video/PDF and a 262,144-token window. $0.10/1M input is not the cheapest. “Leads across the suite” does not hold on every row of Google’s own table
• ECE is best on that chart.

• · Impact: a cheaper, lower-latency control plane for routing and checks.
• Google says Notebook already uses it (up to 40% lower latency on everyday queries).
• Weights are closed; Cloud and prompt caching are not yet available.

【Verifiable】

• ·
• ·
• ·