Google's New Gemini Model Picks Answers Instead of Writing Them

Gemini Decision 1 costs $0.10 per million input tokens, bills nothing for output and already sorts questions inside Gemini Notebook.

Google's newest AI model never writes a word. It looks at a photo, a call recording or a PDF, picks from answers a developer wrote ahead of time and attaches a number for how sure it is.

The model, Gemini Decision 1, went live Wednesday behind a new Gemini Decisions API, Google said in its announcement. Google charges developers $0.10 for each million tokens they feed in, and the answers that come back are free. Developers can prototype for free in Google AI Studio. Support in Gemini on Google Cloud is, in Google's words, "coming soon."

Most people won't call it directly, but it's already working inside Google as a traffic cop. Google says Gemini Notebook uses it to read each incoming question and send it to the right model tier. Everyday questions go to a faster tier and the hard ones go to the deep-reasoning models, which Google says cuts latency on everyday questions by up to 40%.

Search uses it as a safety layer for MCP tools, the add-ons an AI agent can call, flagging adversarial behavior and content. A Google Labs product called CC uses it for something lighter: deciding which of its more than 200 emoji fits as a reaction.

The trick is that the model doesn't build an answer word by word. It scores every option it's given in a single pass. One request can carry up to 1,000 questions against a single input as large as 256K tokens. Those questions can mix yes-or-no checks, multiple choice and graded rubrics. Google puts the end-to-end wait at 150 to 250 milliseconds.

Google's sample code gives the flavor. It shows the model a photo of a breakroom counter along with a note about a 9 a.m. client meeting. Then it asks whether there's a spill, whether someone should grab a cloth or call janitorial, and how bad the mess is. Google's own pitch reaches further: a recorded phone call paired with account data, or a scanned claim form paired with photos of the damage.

Google says the model leads its benchmark suite and has the lowest calibration error. Google's own table, as MarkTechPost laid it out, is less tidy. Decision 1 tops seven of nine accuracy tests. Two of those wins come on audio and video tests where two of its three rivals had no reported scores. It trails TypeSafe's Jev 1.14 on the other two, including When2Call, 79.9% to 82%.

Price is where the post's math gets a closer look. Google calls the model nearly 5x cheaper, but the comparison is to Google's own Gemini 3.5 Flash-Lite. Across the shared test suite, the table puts Decision 1 at $0.88 and Flash-Lite at $4.14. Jev 1.14 ran the same suite for $0.68, and OpenAI's GPT-6 Luna, sold through OpenAI's own Decisions API, ran it for $0.93. By MarkTechPost's reading, Google's model wasn't the cheapest of the four, and these are figures Google reported itself.

The billing has fine print. Ask one question or 1,000 about the same file, and Google bills that file a single time per request. There's no context caching or Batch API yet, MarkTechPost noted, so sending the same long recording twice means paying for it twice. On OpenRouter, the model runs through a separate decisions endpoint, and OpenRouter's model page says standard chat-completions SDKs won't work with it.

Every answer comes back with a confidence score between 0 and 1. Google's advice for the shaky ones is to hand them to a fallback system or a human reviewer. Where that line sits, and how often a person ever sees the claim form, is up to whoever writes the code.

Source: blog.google, retrieved November 12, 2026. Other sources: marktechpost.com, openrouter.ai.