Opera said in its 27 October 2026 announcement that an opt-in WebAI preview is rolling out in Opera One. The feature gives websites and extensions a browser-provided route to models running on the user’s computer. Opera GX and Neon are listed as later releases. For developers, the immediate opportunity is to test local inference without operating a separate model server for those calls. It remains an early preview rather than an established cross-browser platform.
How WebAI handles a request
Opera’s developer explanation describes WebAI as a proposed web standard, with a model-agnostic implementation in its browser. A compatible site or extension can ask the browser’s local model to process text. The published developer material names Prompt, Decision and Semantic Embeddings APIs for generating text, classifying information and finding relationships by meaning. Opera also provides API playgrounds that show availability in the browser being used. The playground catalog should not be read as a promise that every API works on every device.
The local inference path can avoid sending the material being processed to a cloud AI provider, according to Opera. That describes the model call, not every possible network action by a third-party site or extension. Teams still need to inspect an extension’s permissions and wider data handling before treating a workflow as private.
Consent and supported hardware
WebAI is off until the user enables it. Opera says the model download starts only after that choice, and site-specific permissions govern which websites may use local AI. Its developer instructions direct Opera One users to Settings, then Opera AI, to enable the Web AI API. A team evaluating the preview should check the installed browser version, model download, available APIs and permission behavior in its own environment.
Opera lists 64-bit Windows PCs with Intel or AMD processors and Macs with Apple silicon as supported early-preview hardware. Its developer guidance excludes Intel-based Macs, Linux and Windows on Arm. The company says performance depends on the user’s hardware. It has not supplied an independent latency comparison or a guarantee that the same APIs behave identically across browsers in the material reviewed for this article.
What developers should test
A practical evaluation can start with one limited site or extension task, such as classifying page text or generating a short summary. Measure responsiveness, memory use and output quality on the target devices, then check how permission denial or revocation affects the experience. The value proposition is local processing for supported tasks; its actual reliability, resource cost and privacy boundary depend on the model, device and surrounding application. Opera’s preview provides a development path, but those workload-level results remain to be established.
