Google Just Set the Price Floor for AI Decisions: 1,000 Questions Per Pass, 150ms, $0.10 Per Million Tokens
On November 11, 2026 Google launched the Gemini Decisions API, powered by a new Gemini Decision 1 model. One request can score up to 1,000 bounded questions against text, images, audio, video, or PDFs in a single forward pass, returning probabilities and confidence scores in 150-250 ms at $0.10 per million input tokens with zero output tokens billed. Google says it is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API. The agent stack just got a price floor.
Published · By Alfred Belvedere
Today's trend: On November 11, 2026, Google introduced the Gemini Decisions API, a new endpoint on the Gemini API (available through Google AI Studio) powered by a new model called Gemini Decision 1 (gemini-decision-1). Instead of generating answer text token by token, Gemini Decision 1 scores candidate options in a single forward pass and returns probabilities plus confidence scores. The endpoint accepts one shared multimodal input - text, images, audio, video, or PDFs up to 256K tokens, with per-question contexts capped at 16K - alongside up to 1,000 independent questions per request. Three decision types are supported: yes/no with optional true and false criteria, multiple choice from 2 to 255 options returning normalized probabilities for every candidate, and ordered-rubric scoring returning a continuous weighted score. Latency is 150-250 ms end-to-end, with 50-100 ms of that being model compute, and latency stays nearly flat from 1 to 100 questions per request. Price is $0.10 per million input tokens with zero output tokens billed, and Google says that is nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite on the Interactions API. Google claims Gemini Decision 1 leads on agent routing, reasoning, ranking and multimodal understanding with the lowest calibration error, but it published no per-task accuracy numbers for the suite it names. Notably, the model does not generate text at all - your code branches on a number instead of parsing a sentence - and it has no caching, no batch support, and no Vertex AI availability yet as of November 12, so it is limited to the Gemini Developer API (v1beta) in AI Studio. Ecosystem integrations rolled out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with Google Cloud support listed as coming soon. Decision 1 is already running inside Google itself: dynamic model routing in Gemini Notebook, the rapid canvas in Stitch, and Search, which uses the API to detect adversarial behavior on MCP tools. The signal: the lab that owns Flash-Lite just sold a dedicated classification endpoint at a tenth of a cent per million input tokens, which sets a price floor for every judge, router and guardrail vendor in the market. The second-order effect is on agent architecture. For years, routing and scoring were expensive enough that teams collapsed them into a single heavyweight model call - one prompt asking the model to think, judge and answer at once. At a tenth of a cent per million tokens with confidence scores attached to every answer, that trade-off inverts. You can now put a cheap, fast, multimodal decision layer in front of every agent, route ambiguous cases to a human on the confidence score alone, and stop paying reasoning-model prices for what is really a classification problem. The uncomfortable part for incumbents is that the same week MMDM benchmarked 28 decision models and found multi-select accuracy topping out at 37 percent - Google shipped the endpoint without publishing accuracy against a benchmark nobody else has run yet. Cheap decisions with unverified accuracy is a bet, and for the operators building on it, the bet is cheap enough to make and cheap enough to undo.