Chrome Built-in AI in 2026: The Gemini Nano Guide

1. Context: From Origin Trial to Stable

For two years, built-in AI lived behind flags and the Early Preview Program: early samples used a window.ai shape, and every API required an origin-trial token. That era is over for the core set. Starting with Chrome 138, the Summarizer, Translator, and Language Detector APIs are stable, and the Prompt API joined them — first for Chrome Extensions in 138, then for any website in Chrome 148.

One naming note that will save you debugging hours: the current 2026 API surface uses the LanguageModel global — LanguageModel.availability() then LanguageModel.create() — not the window.ai shape from the earliest previews. If you copy a 2024-era snippet and it throws, that rename is almost certainly why.

Under the hood every API runs on Gemini Nano, a compact model Chrome downloads once and shares across origins. Your site never ships the runtime — the browser does — which is why there is no API key and no per-token bill.

2. Four Implications for Developers

Stable on-device inference changes the economics of AI features. Here are the four consequences I consider load-bearing for extension and web developers:

3. The Code: Feature-Detect, Create, Stream

Every built-in AI API follows the same rhythm: check availability() with the same options you will create() with, create a session, then prompt. The snippets below come straight from the official Chrome developer docs (Prompt API, updated Aug 26 2026) — no invented flags, no invented method names. Since Chrome 149 the Prompt API also supports multilingual sessions (en, es, ja, de, fr) and multimodal input (text/image/audio in, text out) plus JSON-schema structured output via responseConstraint.

// Pass the same options to availability() and create().
// expectedInputs: text | image | audio — expectedOutputs: text only.
// Languages (Chrome 149): "en" | "es" | "ja" | "de" | "fr".
const status = await LanguageModel.availability({
  expectedInputs: [{ type: "text", languages: ["en"] }],
  expectedOutputs: [{ type: "text", languages: ["en"] }]
});
// "available" | "downloadable" | "downloading" | "unavailable"

if (status === "available") {
  const session = await LanguageModel.create({
    expectedInputs: [{ type: "text", languages: ["en"] }],
    expectedOutputs: [{ type: "text", languages: ["en"] }]
  });
  const result = await session.prompt("Summarize this in one sentence: ...");
}

// samplingMode ("most-predictable" … "most-creative") is web origin-trial
// only; extensions still use legacy topK/temperature + LanguageModel.params().
// const session = await LanguageModel.create({ samplingMode: "creative" });

const stream = session.promptStreaming("Rewrite this paragraph:");
for await (const chunk of stream) {
  process.stdout.write(chunk);
}

// Structured output (JSON Schema) — same shape for prompt()/promptStreaming():
const schema = { type: "object", properties: { rating: { type: "number" } } };
const raw = await session.prompt("Rate this feedback 0-5 as JSON:", {
  responseConstraint: schema
});
const { rating } = JSON.parse(raw);

const summarizer = await Summarizer.create({
  type: "key-points",
  format: "markdown",
  length: "medium"
});
const summary = await summarizer.summarize(longText);

const translator = await Translator.create({
  sourceLanguage: "en",
  targetLanguage: "es"
});
const text = await translator.translate("Hello, world");

const writer = await Writer.create({
  tone: "formal",
  format: "plain-text",
  length: "medium",
  sharedContext: "Release notes for a Chrome extension"
});
const draft = await writer.write("A new offline mode");

const rewriter = await Rewriter.create({
  tone: "more-casual",
  length: "shorter",
  format: "plain-text"
});
const revised = await rewriter.rewrite(draft);

4. What to Do This Week

A concrete five-step plan to go from zero to a shipped on-device feature before Friday:

5. Sources

All API names, statuses, and code shapes above are taken from official Chrome documentation and the W3C explainer — never from third-party tutorials alone:

Conclusion

Chrome turned the browser into an AI runtime: Gemini Nano ships with the browser, the Prompt API is stable, and the task APIs cover the most common jobs for free. The teams that win in 2026 will treat on-device inference as a new free tier in their stack, not as a replacement for cloud models.

Start with one summarize button behind an availability() check and a cloud fallback. That single pattern — detect, create, stream, degrade gracefully — is the whole game, and it works in production today.