Gemini arrives on CarPlay and boosts its Decisions API

Google deploys Gemini on CarPlay to respond to voice commands while driving, with iPhone synchronization. The company also launches a Decisions API to sort, route, and score multimodal queries in a single pass.

Google deploys Gemini on CarPlay and opens it to all iOS users, while a new Decisions API debuts for developers. In the car, Gemini Live becomes a voice companion that listens, responds, and resumes conversations without leaving the iPhone screen. On the software side, the new API targets bounded decisions — choosing a tool, filtering a policy, ranking a result — with typed responses and a confidence score. Two distinct announcements, but one goal: to move AI from demonstration to practical use.

Dans la voiture, Gemini veut rester utile sans trop distraire

On CarPlay, Gemini displays a simple interface with a waveform indicating when the assistant is listening, thinking, and responding. End and Mute buttons appear at the top left, and the user can interrupt a response to add details, change the subject, or pause. Google says it designed “Gemini on CarPlay to work like a conversational driving companion that minimizes distraction.”

The targeted use is clear: preparing the day during the morning commute, summarizing emails, requesting a brief via Google Workspace, finding a place to eat, or rehearsing a presentation aloud. CarPlay conversations sync with the Gemini app on the iPhone, allowing users to resume or start new exchanges later. Deployment is announced worldwide from November 11, 2026, in all languages compatible with Gemini Live.

Une API pensée pour des choix nets, pas pour rédiger des paragraphes

The Gemini Decisions API, announced by Google on November 11, 2026, processes up to 1,000 independent questions on text, images, audio, video, and PDF in a single request. The principle differs from a classic chatbot: it does not produce explanations but provides actionable decisions for programs. In other words, the API is used to connect, sort, rank, or route, without pretending to philosophize.

Google describes use cases such as tool selection, compliance checks, model routing, context evaluation, or ranking. It is based on Gemini Decision 1, with a context window of 262,144 tokens, yes/no questions, multiple choice, or ordered scales, and a confidence score from 0 to 1 for each answer. Questions are independent, forcing developers to manage dependencies in their own application logic.

Ce que Google met en avant pour les développeurs

Regarding performance, Google indicates a median model compute latency of 50 to 100 milliseconds, and end-to-end latency of 150 to 250 milliseconds on multimodal inputs up to 16K tokens. Latency remains nearly stable between one and 100 questions per request. The company also states that Gemini Decision 1 achieves the best results in its benchmark suite, with the lowest calibration error [to be verified depending on exact test scope].

The announced price is $0.10 per million input tokens, with no charge for output tokens. Shared text cost is counted only once per request, regardless of the number of questions, and Google claims a total cost about five times lower than Gemini 3.5 Flash-Lite with the Interactions API on its reference suite. Access is already available via the Gemini v1beta API, with Python, JavaScript, or REST SDKs, while integrations are underway at OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode.

Pourquoi cette double annonce compte

Taken together, CarPlay and the Decisions API show the same strategy: making Gemini useful where the user already acts, not just in a chat window. In the car, the challenge is contextual voice assistance. In software products, it is about fast, repeatable, measurable decisions where teams want to avoid ambiguous outputs. Google is advancing on two fronts, the consumer interface and the infrastructure component. The kind of tandem product teams like, and that makes security managers sweat a bit.

  • Gemini on CarPlay has been deployed worldwide since November 11, 2026.
  • Up to 1,000 questions per request with Gemini Decisions API.
  • Context window of 262,144 tokens for the Decisions API.
  • Announced price: $0.10 per million input tokens.
  • Announced end-to-end latency: 150 to 250 milliseconds.
  • OpenRouter, LangChain, and Vercel among cited integrations.

By the numbers

  • 1,000 questions — maximum per request in Decisions API, according to Google.
  • 262,144 tokens — announced context window for the Decisions API.
  • $0.10 — price per million input tokens.
  • 50 to 100 ms — median model compute latency announced by Google.
  • 150 to 250 ms — announced end-to-end latency on certain contexts.

To read