Google launches a 'decisions' API that can ask up to 1,000 questions about the same material without writing out answers

The backstory: writing text and making a call are different jobs

The large models most people use answer by writing one word after another. That is a good fit for a poem, a block of code, or an email. But many real tasks are choices, not compositions: Is there a spill in this photo? Should this call go to a human? Should this request go to a cheaper, faster model or to a more expensive reasoning model?

If every one of those calls has to come back as a paragraph, the system is slower and more expensive, and a simple yes or no can get buried in hedging. Downstream software then has to guess what the model actually decided.

Google is trying to treat those bounded decisions as their own problem.

What launched

On November 11, Google opened the Gemini Decisions API in the Gemini API, powered by a new model, Gemini Decision 1. A developer can send one shared input — text, images, audio, video, or PDFs, up to about 256,000 tokens, a rough measure of how much material the model can look at once — along with as many as 1,000 independent questions.

The questions can be mixed. There are yes-or-no checks, multiple-choice items with 2 to 255 options, and rubric scores with an ordered scale. Every answer includes a confidence score from 0 to 1, so software can send shaky cases to a backup rule or a person.

Unlike a typical chatbot model, Google says this system does not write the answer token by token. It scores the options in a single forward pass — one run through the model that produces scores instead of a stream of words.

Google's own performance figures: end-to-end latency of about 150 to 250 milliseconds, with about 50 to 100 milliseconds of model compute; and in contexts up to 16,000 tokens, latency stays nearly flat as the question count goes from 1 to 100. Pricing is $0.10 per million input tokens, with output tokens billed at zero. The company also says the model has the lowest calibration error on its benchmark suite, meaning its confidence lines up better with actually being right, and that it is nearly five times cheaper than Gemini 3.5 Flash-Lite through the usual generative API on that same suite. Those comparisons are Google's, not a third-party audit.

Where Google says it is already in use

Google says teams inside the company are already using the API as a low-latency control layer. Gemini Notebook uses it to route a query to the right model tier, which the company says can cut latency by up to 40% on everyday questions and reserve deeper reasoning for harder ones. A new canvas experiment in the design tool Stitch uses it to generate clickable wireframes and apply voice or tap edits almost immediately. The experimental app CC uses it to pick a reaction from a palette of more than 200 emoji. Google Search uses it, inside a tight time budget, to check whether a tool call looks adversarial.

Those are Google's descriptions of internal use. Outside developers will still need to see whether the same speed and reliability show up in their own products.

What it could mean for you

If "just decide, don't write an essay" interfaces catch on, you may not notice a new chatbot so much as a few ordinary moments getting faster: shorter waits before a support request is routed, quicker blocks on policy-breaking content, or an assistant that switches more cleanly between easy and hard questions.

The limits are clear. This is built for questions whose options are already defined. It will not write an analysis or give an open-ended recommendation. A high confidence score is not a guarantee the answer is right, so a lot still depends on how a product handles the uncertain cases. The price and speed numbers, for now, are the ones Google has published.

Original source published: Google, Nov 12, 2026, 02:00 Beijing time (UTC+8)