Google has launched the Gemini Decisions API, based on a new model, Gemini Decision 1 (gemini-decision-1). Unlike classic language models, this application programming interface (API) does not generate text: it assigns calibrated probabilities to up to 1,000 closed-ended questions, based on a single input. The announced price is $0.10 per million input tokens, with no charge for output tokens.
- Google lance la Gemini Decisions API, qui attribue des probabilités calibrées à 1 000 questions fermées.
- Le prix est de 0,10 $ par million de tokens d’entrée, sans frais pour les tokens de sortie.
- Google affirme que le modèle devance ses concurrents sur 7 des 9 benchmarks de précision.
- Le modèle reste accessible uniquement via l’API hébergée par Google, et il est en retrait sur When2Call.
According to Google’s blog post, the API targets the small intermediate decisions that punctuate the operation of agents, those programs that chain actions autonomously: routing, rule verification, tool selection, or scoring. A request includes a model identifier, a shared input, and a list of questions. Three types are planned: yes/no questions, choice among 2 to 255 options, and scoring on a grid of 2 to 255 levels. Each response comes with a confidence value between 0 and 1, and Google suggests handing low-confidence cases to a fallback solution or a human reviewer.
Billing Designed for Multiple Questions
The model reads text, images, video, audio, and PDFs in a single pass. Google indicates a median compute time of 50 to 100 ms, and 150 to 250 ms end-to-end, for inputs of up to 16K tokens, a token being a unit of text processed by the model. This duration remains almost stable between 1 and 100 questions. The shared input is billed only once per request, regardless of the number of questions. The context window reaches 262,144 tokens, i.e., 256K.
The model is only accessible via the Gemini Developer API (v1beta), hosted by Google: no open weights or self-hosting is offered. It requires the google-genai SDK 3.2 or later, and a free tier allows prototyping in Google AI Studio.
Results Announced by Google
Google compares Gemini Decision 1 to TypeSafe Jev 1.14, GPT-6 Luna on OpenAI Decisions, and Gemini 3.5 Flash-Lite. These figures are provided by Google, and refusals as well as safety blocks are counted as errors. According to the company’s chart, the model outperforms its competitors on 7 out of 9 accuracy benchmarks, with the lowest calibration error. It scores 92.2% on JevBench (public set), versus 90.0% for Jev 1.14. Its calibration error of 0.0250 is, according to Google, about half that of Jev 1.14, which reaches 0.0494.
The model lags on When2Call, with 79.9% versus 82.0% for Jev 1.14, and its lowest score is 58.7% Macro-F1 on Amazon ESCI. On the common suite of seven evaluations, Google announces a total cost of $0.88, versus $0.68 for Jev 1.14 and $4.14 for Gemini 3.5 Flash-Lite. The input price for Jev 1.14 is $0.042 per million tokens. Finally, context caching and the Batch API are not yet supported, meaning repeated context is billed in full for each request. Google also indicates it already uses the tool internally, notably in Gemini Notebook, which uses it to route queries to different model tiers, with latency up to 40% lower for common queries.
