Google's Gemini Decisions API scores up to 1,000 questions in one call at $0.10 per million input tokens. Use it for triage, not writing
Google shipped a model that does not write anything. It only decides, and that is most of what a CRM workflow needs.
What happened
Google introduced the Gemini Decisions API, powered by a new model called Gemini Decision 1. You send one input (text, images, audio, video or PDFs, up to 256K tokens) plus up to 1,000 independent questions. The model scores the options directly instead of generating text, and returns a probability and a confidence score from 0 to 1 for every answer.
There are three question types: yes/no, multiple choice with 2 to 255 options, and scores on an ordered rubric. Google says end-to-end latency is 150 to 250 ms, and pricing is $0.10 per million input tokens with no output tokens billed. The shared input is billed once per request no matter how many questions you ask. Google uses it internally to route Gemini Notebook queries to the right model tier, which it says cut latency by up to 40% on everyday queries.
Simon Willison notes the API is almost identical to the one from startup Jev, and priced the same as OpenAI's GPT-6 Luna decision model. Jev is still cheapest at 4.2 cents per million input tokens. Integrations with OpenRouter, LangChain, Vercel, LiteLLM and others are rolling out.
My take
A lot of AI steps in client workflows are really decisions dressed up as generation. Is this lead a fit? Which pipeline does this inquiry belong in? Is this ticket urgent? Today most of us ask a chat model for JSON, parse it, and hope the format holds.
A decision endpoint removes the parsing and adds something chat models never gave us cleanly: a confidence number. That is what makes routing safe. My rule for these builds:
- Send one record, many questions. A lead's form answers and notes go in once, and fit, intent, budget band and service line all come back in the same call.
- Set a threshold. Above it, the workflow updates the CRM and assigns the owner. Below it, the record goes to a human review queue.
- Log the probabilities next to the record, so you can tune the threshold after a few weeks of real data.
Keep your generative model for the replies and summaries. Let the cheap decider do the sorting.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
