Google's Gemini Decisions API scores 1,000 questions a call at $0.10/M

Google released the Gemini Decisions API on 11 November 2026, with a new model behind it called Gemini Decision 1 (model ID gemini-decision-1). It is available through the Gemini API in Google AI Studio. The model does not write text. You give it one input and a set of closed questions, and it returns an answer and a probability for each.

Google aims it at the bounded choices inside an AI system: which tool to call, which model tier a query needs, whether content breaks a policy, whether a retrieved passage is relevant. Those are the calls many teams make today with a general-purpose model and a prompt that says "answer yes or no".

What Google says it does

• One shared input per request, up to 256K tokens, across text, images, audio, video and PDFs. Google's examples pair a call recording with account metadata, or a scanned claim with damage photos.
• Up to 1,000 independent questions against that input, of three types that can be mixed: noul (yes/no), choice (2 to 255 options, with a normalised probability for every option) and score (an ordered rubric starting at 0, returning a weighted score and a probability per level).
• A confidence score from 0 to 1 on every answer, which Google says is there so code can send ambiguous cases to a fallback or a human reviewer.
• Latency: Google states a median of 50 to 100 ms of model compute and 150 to 250 ms end to end, "across multimodal inputs and contexts up to 16K tokens", staying nearly flat from 1 to 100 questions per request.
• Integrations that Google says are rolling out: OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini on Google Cloud is "coming soon".

What it costs

Google prices the model at $0.10 per 1M input tokens and $0 for output tokens, with a no-cost tier in AI Studio for prototyping. The shared input is billed once per request however many questions you ask; each question adds only its own instruction and criteria tokens. Google says there are no thinking-token charges and no separate transcription or OCR costs.

Google also says the model is "nearly 5x cheaper" across its benchmark suite than Gemini 3.5 Flash-Lite called through the Interactions API. That compares Google with Google. Microsoft listed its Microsoft-Decision-1 at $0.042 per million tokens in October, as this site reported, so Google's is not the cheapest decision model on price alone.

What is not in the announcement

• Independent results. Google says the model "leads across the suite, with the lowest calibration error" on benchmarks covering agent routing, reasoning, ranking and multimodal understanding. Google ran those tests; the figures are in a chart I could not load.
• Latency at full size. The 150 to 250 ms figure is stated for contexts up to 16K tokens and up to 100 questions. Google gives no figure for a 256K-token input or 1,000 questions.
• Rate limits, data-handling terms and a date for Google Cloud. The code sample in the post uses the v1beta API version.
• Measured results from the internal users. Google says Gemini Notebook routes queries with the API and serves everyday queries with "up to 40% lower latency", and that Search uses it as a safety layer for MCP tools "with high efficacy", with no number for the second claim.

The pushback on Hacker News

The Hacker News thread on the announcement (item 50456791) had about 75 points when I read it. The strongest objection is durability: one commenter doubted the endpoint will exist in six months, pointing to Google's record of closing products, and another replied that a model release is not an app and would be succeeded rather than dropped. A second commenter said it costs a little more than other decision models but is useful where the input is audio or video.

The Decoder notes that Google follows the startup Jev, OpenAI, Cloudflare and Microsoft into this category. With that many options, the comparison that counts is on your own data: a few hundred labelled cases, accuracy per class, and whether a stated 0.9 confidence is right nine times in ten.