Opera opened an early preview of WebAI on October 27, giving websites and browser extensions access to AI models running on the user's computer. Its announcement describes browser-managed inference without a separate local server or API key. For developers, that offers a way to build AI features without operating the model service themselves.
The limits matter at launch. Opera calls WebAI a proposed web standard, and its preview supports 64-bit Intel/AMD Windows PCs and Apple silicon Macs. Linux, Intel Macs and Windows on Arm are excluded. This is reporting on Opera's announcement and published playground, rather than a hands-on test of model quality, speed or network behavior.
The browser takes on the model setup
Opera says its implementation works with compatible local models, including open-weight models supplied through Opera and compatible models already on the device. The announcement identifies Prompt, Decision and Semantic Embeddings APIs for generating text, classifying information and finding relationships between content. That puts several common AI tasks behind browser interfaces instead of requiring each extension to connect to a hosted inference endpoint.
The practical attraction is smaller deployment overhead. Opera says developers don't need to run AI infrastructure or pay per request for local inference. Computation still has a cost: it runs on the user's hardware, and Opera warns that response speed depends on the device. A team considering this route needs to test its actual workload on the computers its readers use.
Setup is opt-in. According to Opera's instructions, users update or download Opera One, open Settings, then Opera AI, and enable Web AI API. That downloads a compatible local model and may take a few minutes. Nothing downloads before the feature is enabled, and users can disable it again.
Local processing has a specific privacy benefit
Opera says content processed by the local model doesn't need to go to a cloud AI provider. That can remove an external inference service from an extension's data flow. It doesn't establish what the rest of a website or extension does with the same content. Treat the local-processing claim as a property of this inference path, rather than a privacy audit of every app using it.
The company's Web AI playground provides API availability checks for the visiting browser and links to experimental extensions. One example, Got It?, generates questions about an article on the device. Opera explicitly describes these extensions as examples rather than built-in browser features. We haven't installed them or verified their output.
For a developer deciding whether to experiment, those boundaries give a useful first filter: supported hardware, a user willing to download a model, and a task that tolerates device-dependent speed. A public website also needs a usable path for visitors whose browsers lack the APIs. A proposed standard and an Opera preview don't establish universal browser support.
Watch for broader platform availability and reproducible measurements on ordinary laptops. Opera's launch post acknowledges hardware-dependent performance without supplying latency figures for these tasks. Until those measurements exist, the preview provides a concrete integration route to investigate, while its suitability for a particular product remains a testing question.