Google announced the Gemini Decisions API on November 11, 2026, powered by its new Gemini Decision 1 model. Its job is deliberately narrower than a chatbot’s: score predefined choices against an input, rather than generate an open-ended answer token by token. Google says it returns probabilities and confidence scores in a single pass. Source
Google lists a price of $0.10 per million input tokens, with no output-token charges. That attention-grabbing zero applies to output, not the entire request. Google charges for the shared content’s input tokens once per request, even when several questions use it. Each question’s instructions and criteria add billable input tokens. Source
What it actually decides
According to Google, a shared input can combine text, images, audio, video and PDFs, with support for up to 256K input tokens and 1,000 independent questions per request. Those aren’t 1,000 invitations to write whatever comes to mind. The API supports three bounded formats: Source
- Yes/no: Estimate whether a condition holds.
- Multiple choice: Select among two to 255 candidates, returning the winner and probabilities for every option.
- Rubric scoring: Evaluate an input against an ordered rubric, returning a continuous weighted score and probabilities for each level. Source
Google pitches this for jobs such as selecting tools, routing requests between models, checking policies and grading context. Every answer includes a confidence score between zero and one, which developers can use to send ambiguous cases to another system or a human reviewer. A confidence score is not a correctness certificate. Source
Google reports median model-compute latency of 50–100 milliseconds, or 150–250 milliseconds end to end, across multimodal inputs with contexts up to 16K tokens. It says latency stays nearly flat from one to 100 questions per request. Those figures don’t establish the same speed at the advertised 256K-token or 1,000-question limits. Source
The leaderboard needs its own confidence check
Google says Gemini Decision 1 leads its benchmark suite with the lowest calibration error—a measure of how well confidence matches correctness. Its claim of being nearly five times cheaper is specifically a comparison with Gemini 3.5 Flash-Lite through the Interactions API across their shared benchmark suite, not a blanket discount against every rival. Source
In its November 11 post, Artificial Analysis reports a score of 66 for Gemini Decision 1 on its Decision Index v2.1, behind Drex Max v1.1 and GLiDE v1.2 at 73. It reports strengths in knowledge and reasoning, but weaker results on other evaluations. These are different evaluation scopes; one leaderboard doesn’t cancel the other. Source
The evaluator also reports a cost of $345 per million tasks for Gemini Decision 1, compared with $103 for Drex Max v1.1. That is the cost of its particular benchmark task basket, not a per-token price or a forecast for every application. Source
Google said developers could start through Google AI Studio, including a no-cost tier for prototyping. Its announcement did not specify that tier’s quotas. Ecosystem integrations were rolling out, while support for Gemini in Google Cloud was still in the “coming soon” queue. Source
Comments