Watch: Chrome Decisions API: the prompts, limits, engines and code reviews behind DecisionModel
Chrome's DecisionModel API landed on 6 to 8 October 2026. The exact prompts, limits, engine status and review findings, plus a comparison with Google's MediaPipe DecisionMaker and local tests of EmbeddingGemma 1, EmbeddingGemma 2 and Laya.
Transcript
Google Chrome is testing a new on-device tool called the Decisions Application Programming Interface, or API. It is designed to let websites ask closed-ended questions, like yes-or-no queries, multiple-choice options, or ratings, about a piece of text, right inside the visitor’s browser.
Instead of sending data to a cloud server, the browser runs a compact artificial intelligence model locally on the user's computer. It then calculates the probability for each possible answer.
However, public code reviews and early tests show the technology is still in its infancy. For one, running these models directly on a computer’s central processing unit can be slow. Processing a few thousand words and a handful of questions can take nearly ten seconds.
Developers also discovered some quirky bugs. The models sometimes show a strong bias toward choosing the first option, letter A, regardless of the text. There are also concerns that a visitor's text could blend into the developer's questions, confusing the system.
Meanwhile, Google has quietly released a sister tool called MediaPipe DecisionMaker. In independent tests, this alternative library was faster and delivered much more accurate results than the experimental Chrome API.
For now, the Chrome Decisions API remains a work in progress, hidden behind developer flags as engineers try to balance privacy, speed, and accuracy on local devices.