Opera WebAI Explained: What Local AI in the Browser Changes

By Jaspal Singh October 27, 2026 Updated: October 27, 2026

Opera's WebAI, released as an opt-in preview in Opera One on October 27, 2026, lets websites and browser extensions call an AI model that runs on your own computer instead of a cloud server. Text you feed those features can stay on the device, developers skip per-request API bills, and your hardware does the work.

What Happened

On October 27, 2026, Opera announced WebAI in a post on its official blog. The company describes it as a proposed web standard and says it is backed by companies including Google and Microsoft. The feature ships first as an early preview in Opera One, the company's main desktop browser. TechCrunch and 9to5Mac both reported that Opera GX, the gaming browser, and Opera Neon, the agent-focused browser, are due to follow, though Opera has not given a date.

According to Opera's post, the preview runs on 64-bit Windows PCs with Intel or AMD processors and on Macs with Apple silicon. It is not available on Intel-based Macs, Linux, or Windows on Arm. At launch the browser can use Google's open-weight Gemma models, and on compatible Macs it can also use Apple's on-device foundation model, the same small model Apple ships for its own on-device features.

Opera's post names three developer interfaces: a Prompt API for generating text, a Decision API for classifying information, and a Semantic Embeddings API for measuring how closely pieces of content relate. Opera's developer playground at next.opera.com lists a wider set of eight: Prompt, Summarizer, Writer, Rewriter, Translator, Proofreader, Semantic Embedder and Decisions. Opera also published six experimental extensions built on WebAI, including a tab organiser and a clickbait flagger.

This is not Opera's first experiment with local models. In April 2024 it added access to locally run language models in a developer build of Opera One, which it calls a first for a major browser. Krystian Kolondra, Opera's EVP of browsers and AI, framed the new release as taking that work from developers to everyone, adding that "a lot of what people want from AI doesn't need a data center."

Why It Matters

Most AI features on websites today work the same way: the page sends your text to a server run by OpenAI, Google, Anthropic or another provider, the provider runs a large model, and the answer comes back. That model has three costs. The site pays for every request, you need a network connection, and your text leaves your machine and lands in someone else's logs.

WebAI moves a slice of that work into the browser. A site that wants to summarise a comment thread, or an extension that sorts your tabs by topic, can ask the browser's local model instead of a cloud API. The site runs no AI servers, holds no API key, and the text does not have to leave your computer.

For developers, the bigger change is the word "standard." Running a model in a browser tab is already possible with libraries that download weights into the page and run them with WebGPU, but every site then downloads its own copy of a multi-gigabyte model. A browser-level API means the model is downloaded once and shared by every site you allow, and the developer writes a few lines of JavaScript instead of shipping an inference engine.

For users, the trade is simple to state. You get AI features without a subscription or an account, and with more control over where your data goes. In return your laptop does the computing, which means fan noise, battery drain and slower answers on older machines, and the local model is far smaller and less capable than the frontier models behind ChatGPT or Claude.

How It Works

WebAI is switched off by default. To turn it on in Opera One, you open Settings, go to the Opera AI section and enable the option labelled "Enable Web AI API." Opera says this downloads a compatible local model, which can take a few minutes, and you can switch the feature off again at any time. Nothing is downloaded before you opt in.

Once the model is installed, access is gated per site. Digital Trends reports that each website needs your permission before it can use the browser's local AI. That matters because the caller is not only Opera: any page you grant access to can send text to the model on your machine and read back what it produces.

The developer side looks familiar to anyone who has followed Chrome's built-in AI work. Opera's Prompt playground points developers to Google's Prompt API documentation as its reference, and the names of Opera's eight playground APIs match the built-in AI APIs Google has been shipping in Chrome. That indicates Opera is implementing a shared interface rather than inventing its own. In that interface, a page first calls LanguageModel.availability() to learn whether a model is ready, still downloading or unavailable. It then calls LanguageModel.create() to open a session and sends text with session.prompt(), or with session.promptStreaming() to receive the answer word by word. Sites choose a sampling mode, such as balanced or creative, rather than tuning raw model settings, so the same code behaves sensibly whichever model sits underneath.

The Prompt API itself is developed in the open by the W3C's Web Machine Learning Community Group, whose explainer and draft specification are public on GitHub. The explainer says experimental implementations exist in Chrome and Edge, which fits Opera's claim of Google and Microsoft backing. It also spells out several design choices that shape what WebAI can promise:

Model-agnostic by design. The specification does not say which model a browser must use. Chrome uses Gemini Nano; Opera uses Gemma or Apple's model. A browser is even allowed to ship no model at all and simply report that none is available.
No quality guarantee. The authors treat output quality and cross-browser consistency as up to each browser. The same prompt can give different answers in Chrome and Opera.
Limited reach inside pages. By default the API is open to top-level pages and same-origin frames only. A cross-origin iframe, such as an embedded ad, needs explicit permission through the language-model permissions policy, and the API is not exposed to web workers.

The other two APIs Opera highlights cover tasks that do not need free-form text. A Decision API takes content and a set of labels and picks the one that fits, which is how an extension can flag a headline as clickbait. A Semantic Embeddings API turns text into vectors, lists of numbers that place similar meanings close together, so an extension can group related tabs or find the passage on a page that answers a question without generating any new text at all. Both are cheaper to run locally than full text generation.

Hardware is the practical limit. Opera says only that performance depends on your device. Google's documentation for Chrome's version of the Prompt API is more specific: it asks for at least 22 GB of free space on the drive holding the browser profile, and either a GPU with more than 4 GB of video memory or 16 GB of RAM with four or more CPU cores. Opera has not published equivalent figures.

What's Still Unknown

Where the standard actually stands. Opera calls WebAI a proposed standard. The Prompt API is a Community Group proposal, an early incubation stage at the W3C, not a finished standard on a formal track. The explainer notes that Mozilla and Apple's WebKit team have open standards-position requests, and neither Firefox nor Safari has committed to shipping it. Opera has not said whether its Decision API is part of that shared work or an Opera addition.

Model sizes and minimum hardware. Opera has not published download sizes, which Gemma variant it installs, or minimum memory figures for its own implementation.

What sites do with the output. Local processing keeps the input away from a cloud AI provider, but a website can still send the model's answer, or anything else on the page, to its own servers. As CyberInsider pointed out in its coverage, WebAI is not a general guarantee about how a site or extension treats your data.

Timing. There is no date for Opera GX or Opera Neon, no word on when the preview becomes a stable feature, and no mobile plan.

Frequently Asked Questions

What is Opera WebAI?

WebAI is a feature in Opera's browsers that lets websites and extensions use an AI model running on your own computer. Opera launched it as an opt-in early preview in Opera One on October 27, 2026, and describes it as a proposed web standard. It supports text generation, classification and semantic search through browser APIs, without a cloud account, subscription or API key.

How do I turn on WebAI in Opera One?

Open Opera One's Settings, go to the Opera AI section and switch on the option called "Enable Web AI API." The browser then downloads a compatible local model, which Opera says can take a few minutes. You can turn the setting off whenever you like. After that, each website still has to ask for your permission before it can use the model.

Which computers can run WebAI?

Opera's preview supports 64-bit Windows PCs with Intel or AMD processors and Macs with Apple silicon. It does not run on Intel-based Macs, Linux or Windows on Arm. Opera has not published minimum memory figures, but says speed depends on your hardware. Chrome's comparable feature asks for 22 GB of free storage plus a capable GPU or 16 GB of RAM.

Which AI models does WebAI use?

At launch, Opera One can use Google's open-weight Gemma models on supported Windows PCs and Macs. On compatible Macs it can also use Apple's on-device foundation model. Opera calls its implementation model-agnostic, and the underlying Prompt API proposal does not require any particular model, so the list could grow as the preview develops.

Does WebAI keep my data private?

It keeps the text you process away from cloud AI providers, because the model runs on your machine. It does not control what a website does afterwards: a site could still send the model's output or other page data to its own servers. Per-site permissions let you decide which sites may use the local model at all.

Is WebAI the same as Chrome's built-in AI?

They appear to share the same developer interface. Opera's Prompt playground links to Google's Prompt API documentation, and its API names match Chrome's built-in AI APIs. The models differ: Chrome uses Gemini Nano, while Opera uses Gemma or Apple's model. Results for the same prompt can therefore differ between the two browsers.

Will WebAI work in Firefox or Safari?

Not yet. The Prompt API is a W3C Community Group proposal with experimental versions in Chrome and Edge, and now Opera. Mozilla and Apple's WebKit team have open requests to state their positions, and neither has committed to shipping it. Until they do, developers need a fallback for browsers without a local model.

Do developers have to pay to use WebAI?

No per-request fees apply, because inference runs on the visitor's computer rather than a paid cloud service. Developers do not need an API key or AI servers. The costs move elsewhere: features must handle visitors whose devices cannot run the model, and output quality from a small local model is lower than from large cloud models.