Browser built in AI compared with your visitors' own AI
Chrome's and Edge's built in models, Firefox, Safari and WebLLM compared with AI Connect on cost, hardware, context, tool calling, phones and release status.
Published on
Built in AI is a language model that runs on the visitor's own device inside the browser, downloaded once and called from a page through a JavaScript API. The site pays nothing per token, and on Chrome no prompt leaves the device. AI Connect brings the site's inference bill to $0 through another path. The visitor's own AI account answers through AI Wallet, the extension that holds their accounts, and the provider runs the model. Both are forms of bring your own AI: the visitor's device or the visitor's account pays for the answer.
Four ways to run a model, on eight points
| Point | Chrome Prompt API | Edge Prompt API | WebLLM | AI Connect |
|---|---|---|---|---|
| Who pays | Nothing per token; the visitor's device runs the model | Nothing per token; the visitor's device runs the model | Nothing per token; the visitor's device runs the model | The visitor's plan, credits or API key |
| Model | Gemini Nano | Phi-4-mini; Aion-1.0-Instruct as a prerelease from Edge 150.0.4070 | Open models such as Llama 3.2 and Qwen3 | The visitor's account and model, out of 20 kinds of account |
| What the visitor needs | Windows 10 or 11, macOS 13 or later, Linux or Chromebook Plus; 22 GB free disk; over 4 GB of video memory, or 16 GB of RAM and 4 cores | Edge Canary or Dev with a flag on; Windows 10 or 11 or macOS 13.3 or later; 20 GB free disk; 5.5 GB of video memory for Phi-4-mini | A browser with WebGPU; 879 MB of video memory for Llama-3.2-1B up to 5,696 MB for Qwen3-8B | Chrome 140 or later on a computer, AI Wallet and an AI account |
| Context | No figure in the docs; an input quota of 6,144 tokens measured in Chrome 137 | 9,216 tokens | 4,096 tokens by default | The context of the visitor's model |
| Tools and MCP | No tool option in the docs; JSON Schema output | No tool option in the docs; JSON Schema or regular expression output | Preliminary tools and tool_choice, marked work in progress | Function tools from the site's MCP server |
| Phones | Android and iOS unsupported | Windows and macOS only | Not stated | No, Chrome on a computer only |
| Offline | Yes, after the first download | Yes, after the first download | Model kept in the browser's cache after the first download | No; the provider runs the model |
| Status on 12 November 2026 | Stable on the web from Chrome 148 | Developer preview from 138.0.3309.2 | Version 0.2.88 of 10 November 2026 | 0.1.0 on npm since 16 November 2026; wallet in the Chrome Web Store's review |
Chrome's Prompt API and task APIs
Chrome ships Gemini Nano, Google's on device model, behind a set of web APIs. The Translator, Language Detector and Summarizer APIs are stable since Chrome 138. The Prompt API, which takes free form prompts, is stable on the web from Chrome 148 and in extensions from 138. Writer, Rewriter and Proofreader run as trials: the Writer origin trial covers Chrome 137 to 148 and the Proofreader trial 141 to 145.
The model needs a computer on Windows 10 or 11, macOS 13 or later, Linux or a Chromebook Plus, with at least 22 GB of free disk. Chrome removes the model when free space falls below 10 GB. The device needs a graphics card with more than 4 GB of video memory, or 16 GB of RAM and 4 cores. The Prompt API takes text, images and audio, returns text, and supports English, Japanese, Spanish, German and French.
Chrome sends no prompt data to Google or any third party when a page uses the model, and the network serves the first download only. Chrome's documentation gives no context size and points to chrome://on-device-internals. In Chrome 137 the download measured 1.5 to 2.4 GB and the input quota 6,144 tokens.
Edge's developer preview
Edge runs Phi-4-mini behind the same Prompt API shape, as a developer preview in its Canary and Dev channels from version 138.0.3309.2. The visitor turns the API on in edge://flags, so a site's ordinary Edge visitors on the stable channel get no model. Phi-4-mini needs Windows 10 or 11 or macOS 13.3 or later, 20 GB of free disk and 5.5 GB of video memory. From Edge 150.0.4070, the prerelease Aion-1.0-Instruct model runs on devices with weaker graphics or on the processor alone. The context is fixed at 9,216 tokens.
Firefox and Safari
Firefox gives web pages no model. Its browser.trial.ml API, built on Transformers.js and ONNX, is experimental and open to extensions only. Safari gives web pages no AI API either: Apple's Foundation Models framework serves native Swift apps.
WebLLM in the page
WebLLM is an open source library under the Apache 2.0 licence that downloads an open model into the page and runs it on the graphics card through WebGPU. The site picks the model, and the visitor downloads it on the first run. The browser's Cache API keeps it for later visits. Video memory ranges from 879 MB for Llama-3.2-1B and 2,264 MB for Llama-3.2-3B to 5,001 MB for Llama-3.1-8B and 5,696 MB for Qwen3-8B. The default context is 4,096 tokens, and tool calling through tools and tool_choice is preliminary.
Which feature belongs on which model?
Built in AI fits a site better for short tasks on capable desktops:
- Offline summaries and translation. Chrome's Summarizer, Translator and Language Detector are stable since Chrome 138 and run without a network after the first download.
- Text that stays on the device. Chrome sends no prompt data to Google or any third party, and the visitor needs no AI account.
- Prompts under the quota. A plain retrieval query averaged 3,865 input tokens in one study, inside Edge's 9,216 and Chrome's measured 6,144.
AI Connect fits a site better when the feature needs the site's data or a larger model:
- Tools. The site's search, catalogue or booking runs as an MCP server (Model Context Protocol, the open protocol for serving tools to a model), and the visitor's model calls it through the SDK. The MCP guide shows the loop.
- Long prompts. An iterative retrieval query averaged 12,814 input tokens, above the windows of Chrome (6,144 tokens), Edge (9,216) and WebLLM (4,096).
- Hardware below the bar. A computer with less than 22 GB of free disk, or with neither 4 GB of video memory nor 16 GB of RAM, gets no Gemini Nano. The visitor's account answers on the provider's servers.
- More accounts. The wallet answers with any of 20 kinds of account, listed on the providers page.
On phones, neither built in AI nor AI Wallet runs. Chrome's built in models do not run on Android or iOS, and AI Wallet runs in Chrome on a computer. On 16 November 2026 a phone visitor answers through AI Connect after a sign in with OpenRouter or the TensorPass plan, and a site that answers phones without a sign in pays on its own key, as the comparison of paying for your own inference sets out.
WebMCP and the browser's agent
WebMCP lets a site publish tools in the page for the browser's own agent, Gemini in Chrome, to call. It entered an origin trial in Chrome 149. The visitor then talks to the browser's assistant, and the site supplies the tools without running a chat of its own. With AI Connect the chat stays in the site's page, under the site's interface and guards. The account the visitor chose for the site answers it.
Combine both in one page
A page calls Chrome's Summarizer for an offline summary of an article and AI Connect for a chat that searches the site's catalogue. The two use separate APIs. The summary needs Chrome's hardware bar, and the chat needs AI Wallet or a sign in with OpenRouter or the TensorPass plan. AI Connect 0.1.0 is on npm since 16 November 2026.
Sources
- Chrome for Developers, Built in AI APIs, read 12 November 2026
- Chrome for Developers, The Prompt API, updated 21 October 2026
- swyx, Gemini Nano in Chrome, 8 July 2025
- Microsoft Learn, Prompt a built in language model with the Prompt API, updated 19 October 2026
- Microsoft Edge Explainers, issue 1224 on the Prompt API context size, 8 January 2026
- LogRocket, Browser AI APIs in 2026, 5 November 2026
- Apple Developer, Foundation Models framework, read 12 November 2026
- MLC AI, WebLLM repository and src/config.ts, version 0.2.88 of 10 November 2026
- Chrome for Developers, Chrome at I/O 26, 19 May 2026
- arXiv 2409.02361, token use of plain and iterative RAG, 2024
- TensorPass, Providers, 12 November 2026
- TensorPass, Changelog, 0.1.0, 16 November 2026
Questions
- Does Chrome's built in AI cost the site anything per answer?
- No. Chrome runs Gemini Nano on the visitor's device and charges nothing per token. The visitor's computer needs 22 GB of free disk and either more than 4 GB of video memory or 16 GB of RAM with 4 cores.
- Do built in browser models call a site's tools?
- The Prompt API documentation of Chrome and of Edge lists no tool option, and both constrain output with a JSON Schema. WebLLM marks its support for tools and tool_choice as preliminary work in progress. AI Connect sends the function tools of the site's MCP server to the visitor's model and calls them from the page.
- Does built in AI work on phones?
- No. Chrome lists Android and iOS as unsupported for its built in models, and Safari gives web pages no AI API. AI Wallet runs in Chrome on a computer only. On 16 November 2026 a phone visitor answers through AI Connect after a sign in with OpenRouter, or with the TensorPass plan on a site that passes a site key, and a site that answers phones without a sign in needs its own key.
- Which browsers give web pages a built in model on 12 November 2026?
- Chrome, where the Prompt API is stable on the web from version 148. Edge offers it as a developer preview in its Canary and Dev channels behind a flag, Firefox exposes its experimental model API to extensions only, and Safari exposes none.