Google Starts Selling a Gemini Model That Writes No Text and Bills No Output

Google said its new Gemini Decision 1 model answers yes/no, multiple-choice and rubric questions in a single pass, billing $0.10 per million input tokens and nothing for output, through a new Gemini Decisions API.

Tool selection, policy checks, model routing and context grading are the bounded choices Google built the endpoint for, and a chat model answers them by generating text. Decision 1 prices that step on input alone, so the cost of a decision depends on how much the model reads, not on how much it writes. According to a published report, Jev 1.14 posted the lower total in Google’s own comparison, so a buyer choosing a model for routing calls has to weigh what each decision costs against what each model can read.

How is it billed?

One request carries a single input, text, images, audio, video or PDFs up to 262,144 tokens, and up to 1,000 independent questions about it. The model returns a probability and a confidence score for each answer in 150 to 250 milliseconds end to end, Google said. The shared input is billed once however many questions follow, and each question adds only its own short instruction and answer options to the bill. Audio and scanned pages carry no separate transcription or text-extraction fee.

Across the benchmark suite Google ran, the model costs roughly one-fifth as much as Gemini 3.5 Flash-Lite run through the Interactions API.

Who costs less?

Jev 1.14, from TypeSafe, does. On the seven tests that both models ran, Google’s own table puts the total at $0.68 for Jev and $0.88 for Gemini Decision 1, the report said. By The Denominator’s arithmetic, Google’s model costs about 29% more. It said Jev also beat Google’s model on two of those tests, Mind2Web Multimodal and When2Call, while Google’s model led on the other five and showed the lower calibration error, 0.0250 against 0.0494. Calibration error measures how far a model’s stated probabilities sit from how often its answers are right.

Google ran every one of these tests itself. The two audio and video tests have no Jev score.

Where can a team use it?

The model is available through the Gemini API in Google AI Studio, with a no-cost tier for prototyping. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, Google said, and it gave no date for support in Gemini in Google Cloud. Inside Google, Gemini Notebook already uses the endpoint to route queries between model tiers, with up to 40% lower latency on everyday queries.

How many of the company’s agent calls only route or check, and what would they cost with every output token free?