Opera is betting that a large share of everyday artificial intelligence work does not need to happen in a distant data center at all. The browser maker has begun previewing a feature called WebAI, which allows websites and browser extensions to run AI models locally inside Opera One, using the computing power of the machine already sitting on the user's desk. No separate AI subscription is required, and the information being processed does not have to leave the device.
The pitch is deliberately unglamorous. Instead of pitching chatbots that write poetry or generate video, Opera is focused on small, repetitive chores: sorting a chaotic pile of open tabs, classifying blocks of text, flagging clickbait before it swallows another five minutes of the day. Those tasks, the company argues, are exactly the kind of work that should never have required a round trip to a server farm in the first place.
How WebAI handles a request
WebAI is best understood as a bridge between a web page and a model that already lives on the computer. A website or an extension asks the browser for access to local inference. If the user has switched the feature on and granted that particular site permission, the request is routed to a model running on the device rather than to an outside API. The browser handles the plumbing, the site handles the interface, and the model handles the actual computation.
Opera describes its implementation as model-agnostic and centered on open models. On Opera One, the browser can call on Google's Gemma family of models. On compatible Macs, it can also reach Apple's on-device foundation model, which is already built into recent versions of macOS for Apple Intelligence features. That flexibility matters because it means the same web-facing code can run against different engines depending on what hardware and operating system the visitor happens to be using.
A browser standard, not just a product feature
The important framing here is that Opera is not simply shipping a proprietary feature. It is positioning WebAI as a proposed web standard, and it says the effort has backing from Google and Microsoft, the two companies whose browser engines and platform decisions largely determine what the web can and cannot do. If a standard like this takes hold, a developer could write one integration and have it work across browsers, rather than building a separate pipeline for each vendor's AI stack.
That ambition fits a broader shift across the browser industry. Over the past two years, the major browser makers have all moved toward putting small language models directly on user devices. The reasoning is consistent across vendors: on-device models are cheaper to operate at scale, respond with less latency because there is no network hop, and keep sensitive text on the machine. Opera's contribution is to suggest that those local models should be exposed to the open web through a permissioned browser API, rather than being locked inside each vendor's own assistant or sidebar.
What developers stop paying for
The economics for developers are straightforward and probably the most persuasive part of the proposal. Running AI through a hosted API means paying for every single request, and those costs scale linearly with usage. A popular extension that classifies thousands of snippets a day can turn into a real monthly bill, and a small independent developer may not be able to absorb it at all.
With local inference through the browser, the marginal cost of a request drops toward zero from the developer's perspective. There is no server to maintain, no API key to rotate, no rate limit to design around, and no bill that spikes when a project goes viral. Opera's own examples reflect that freedom: one extension can reorganize a messy tab collection, and another can scan a page for clickbait-style headlines. Neither needs a backend, and neither needs a subscription to function.
Permissions, opt-ins, and the privacy trade
Opera has kept WebAI disabled by default, which is the right instinct for a feature that lets third-party code reach models running on a user's own hardware. Nothing is downloaded until the user explicitly enables the capability, and each individual site must request permission before it can call the browser's local AI.
That layered consent model matters more than it might first appear. When AI runs inside a single company's assistant, there is one relationship to reason about. When any website can request local inference, the trust boundary becomes much thinner. A site with access to a general-purpose language model running on a user's laptop has a capability that did not exist in the browser a few years ago, and the permission prompt is the only thing standing between a curious page and that capability.
The hardware caveat nobody can design around
There is a catch, and it is a physical one. No data center will rescue weaker hardware. When inference runs locally, speed and capability depend entirely on the computer doing the work. A recent laptop with a capable neural processing unit will handle small models comfortably; an older or budget machine may struggle, throttle, or drain its battery doing the same job.
Opera has not disclosed minimum performance requirements beyond listing supported platforms, which leaves an open question about how the experience degrades on modest hardware. There is also no graceful fallback to a cloud service built into the design as described, which means a slow device simply gets slow results. HTML-in-Canvas and other experiments hint that Opera is thinking about the browser as an application platform, but local AI will always be bounded by the silicon underneath it.
The bulldozer next door
Opera is pairing WebAI with a separate experiment called HTML-in-Canvas, and the pairing says something about where the company thinks browsers are heading. The feature lets developers place fully functional webpages inside 2D and 3D scenes, with buttons and forms that keep working even as the page moves through virtual space.
The showcase includes a Gamer Room with live webpages playing on its screens, and a Webpage Bulldozer that turns site elements into objects a user can shove around. Both demos already work in Opera browsers without any setup, which means the bulldozer is the part anyone can try today, while WebAI remains an opt-in early preview that requires users to deliberately turn it on.
Opera GX, the company's gaming-focused browser, and Opera Neon are both scheduled to receive WebAI at a later date.
Source:Digital Trends News