Opera Pushes WebAI Standard to Run AI Models in Browser

Opera released WebAI on Tuesday, a proposed web standard that lets websites and browser extensions run AI models on the user's own machine instead of a vendor's cloud. The feature is live as an early preview in Opera One, with Opera GX and Opera Neon to follow.

What Opera actually shipped

WebAI is not a model and not an app - it is an interface. Opera says the standard is backed by companies including Google and Microsoft, and its implementation exposes three APIs: Prompt for text generation, Decision for classification, and Semantic Embeddings for finding relationships between pieces of content.

Inference runs inside the browser itself. There is no separate AI application to install, no local server to configure and no API key to obtain. "Developers can add AI features without running AI infrastructure or paying for every request," Opera said in its announcement. Compatible models include open-weight systems such as Google's Gemma, plus Apple's on-device foundation model on Macs.

Opt-in is explicit: nothing downloads until a user opens Settings, goes to Opera AI and enables the Web AI API. The preview works on 64-bit Windows PCs with Intel or AMD chips and on Apple silicon Macs - not on Intel-based Macs, Linux or Windows on Arm.

"Your laptop can already do it, without sending your data anywhere," said Krystian Kolondra, Opera's EVP of browsers and AI.

Why builders should care

The economics are the interesting part. A per-request cloud bill becomes a one-time download, and a privacy-policy argument becomes a local processing claim. For small teams shipping a single-purpose feature - summarizing a page, tagging saved articles, ranking results - that removes both the inference invoice and most of the data-handling burden.

Opera shipped six experimental extensions to demonstrate the pattern, including Project Lens, which shades an article by how relevant each paragraph is to a stated project. Developer documentation sits at next.opera.com/web-ai. The company has been here before: in 2024 it added experimental support for roughly 150 local model variants in a developer build. What changed is consumer hardware and open-weight model quality.

The trade-offs

Local inference spends the user's battery and thermal budget, not a data center's. Opera concedes that response speed depends entirely on the device and that this is a preview. A site tuned to sub-second latency on an Apple laptop will feel different on a mid-range Windows machine.

Distribution is the other limit. A web standard needs more than one browser shipping it, and Opera's share is small. Until Chrome, Edge or Safari implement the same surface, WebAI is a differentiated feature rather than a platform a product team can depend on.

What to do now

Build the extension, not the business. WebAI fits personal tooling and prototypes where on-device processing is the point. Keep a cloud fallback behind a feature flag, and watch whether Google and Microsoft's backing turns into implementations.