Google DeepMind announced the Gemini Decisions API on November 11, 2026. The service is built around Gemini Decision 1, a model that can answer up to 1,000 independent yes/no, multiple‑choice or rating questions in one go. It works on text, images, audio, video and PDF files. The company says the model returns probabilities and confidence scores in 150 to 250 milliseconds. The API is already used inside Google. It helps route tasks in Gemini Notebook, guides design ideas in the Stitch prototyping tool, and checks safety in Google Search. For developers, the API is live through Google AI Studio and will soon appear in integrations such as OpenRouter, LangChain and Vercel. Gemini Decision 1 also scored a kitchen‑counter photo in a single pass, flagging a stain (yes/no confidence 0.95), picking the right cleaning action (98 % confidence) and rating severity on a 0‑2 scale (1.19). Google says the model leads benchmark tests for agent routing, reasoning and multimodal understanding, and it recorded the lowest calibration error among peers.
How does Gemini Decision 1 compare to other AI services?
The pricing is $0.10 per million input tokens with no charge for output tokens. Google notes this is about five times cheaper than its Gemini 3.5 Flash‑Lite model. By contrast, earlier decision‑type models from startups like Jev and larger providers such as OpenAI, Cloudflare and Microsoft have appeared only in the past few weeks, but pricing details were not disclosed. The speed advantage—answers in a few hundred milliseconds—means the API can be used for real‑time applications like interactive assistants or rapid safety checks. However, the model only handles closed‑ended formats; it does not generate free‑form text or long‑form explanations.
Why it matters
For developers, the new API offers a cheap, fast way to add large‑scale decision‑making to apps, from quick safety filters to real‑time survey scoring. For end users, it could mean smoother experiences in products that need rapid yes/no or multiple‑choice answers, though the model currently works only with closed‑ended formats.
