Opera’s WebAI preview moves model execution into the browser—but privacy is still a workflow question

What is available in the preview

Opera’s October 27 announcement introduces WebAI as an opt-in early preview in Opera One. The supported hardware is specified: 64-bit Windows PCs using Intel or AMD processors, and Macs with Apple silicon. Opera GX and Opera Neon are named as future destinations, not browsers confirmed here to offer the same preview today.

Models are not downloaded until the user activates the capability, according to the official preview announcement. Site-specific permissions let the user choose which websites can access local AI. The feature should therefore not be described as automatically enabled or as preinstalled models available to every website.

TechCrunch separately reports the introduction of Opera’s local-AI proposal. The announcement also introduces HTML-in-Canvas, for placing live web content inside three-dimensional environments. That is a separate capability, not another name for WebAI or evidence that the local-model feature supports every web environment.

The release links this work to Opera’s experimental developer build offering local models in 2024. That is historical context from the vendor, not the event being presented as new. The October 27 development is the specified WebAI preview and its user-access controls, not a claim that browser-based local inference has never existed before.

Device-side inference versus a remote service

In the device-side arrangement Opera describes, the local model uses the user’s computer rather than a remote model endpoint for its task. Opera says developers can add suitable AI capabilities without operating inference infrastructure or paying for each request. These are the vendor’s proposed advantages; Foundra has not tested a feature, device or workload.

A remote service and a device-side model place operating responsibilities in different places. Remote inference depends on the supplier and service arrangement. Local inference depends on the user’s supported hardware and the browser capability. Opera explicitly says performance depends on the device. A claim that local execution is universally faster or cheaper would require a comparison the announcement does not provide.

The preview’s hardware restriction is therefore part of the product decision, not a footnote. A startup cannot infer support for all of its users from support for specified PCs and Macs. It would need a deliberate path for users outside the documented preview, and it should not count the planned GX and Neon rollout as already delivered reach.

Opera’s examples include sorting tabs, identifying relevant job posts and flagging clickbait. They illustrate the kinds of device-side tasks the company envisages; they are not independently demonstrated accuracy results. A narrow task may fit this arrangement well without establishing that every AI feature should move away from a remote service.

Permission is a control, not a whole-workflow guarantee

The opt-in activation and site-specific permissions are substantive controls described by Opera. They identify two user choices: whether to obtain the capability and which sites may use it. The privacy comparison should recognise those controls rather than suggesting that the release offers no account of access.

The release’s local-inference claim still has a defined scope. Keeping information from being sent to a remote AI service for this task is different from establishing what the surrounding website or extension does with information elsewhere in its workflow. That is an architectural distinction, not an allegation that Opera violates its announced local-model boundary.

A broader assessment would follow the complete information path, including what the site receives and retains and which other operations use the network. The available permission controls help define model access, but they do not independently audit every application using that access. No whole-browser privacy certification is inferred here.

There is also a standards boundary. A vendor’s introduction of a proposal does not establish universal adoption across browser engines. Developers would need current support and documentation for the intended audience before relying on WebAI as a general web capability. The supplied evidence supports an early preview with specified controls and hardware, not a settled cross-browser standard.

Opera’s launch is useful as another concrete architecture to evaluate. Its comparison with cloud inference is not a single winner-and-loser decision: it is a choice about task size, device support, permission and operating responsibility. The documented preview makes those boundaries more specific, while leaving performance, broader compatibility and whole-workflow privacy to be evaluated on their own evidence.