Opera WebAI: Local AI APIs, Supported Devices and Preview Limits

Opera announced a WebAI early-preview rollout on October 27, letting websites and extensions use models running on the user's computer. The practical change is that developers can build some AI features without sending each request to a cloud model or paying a token-based inference bill.

The launch announcement names Opera One as the first browser, with Opera GX and Opera Neon to follow. It is an opt-in preview, not a universally available browser capability.

What Opera WebAI actually provides

WebAI gives web applications access to AI through the browser. Opera describes it as a proposed web standard, so it should not be treated as a finalized standard implemented consistently across browsers.

Opera's implementation can use compatible local models, including Google's Gemma and Apple's on-device foundation model on Macs. Users enable it before a model is downloaded, and site-specific permissions control which websites can access local AI. Opera says no separate AI subscription is required.

This changes where a feature's computation happens. The website supplies the experience, while the user's device supplies the inference resources. That can make small, frequent operations easier to offer, particularly when charging for every classification would make a lightweight feature awkward to price.

Text, decisions and semantic search

The WebAI playground lists eight API categories: Prompt, Summarizer, Writer, Rewriter, Translator, Proofreader, Semantic Embedder and Decisions. It checks availability in the browser that opens it rather than promising every API works everywhere.

The distinctions matter. A prompt interface produces language. An embedding represents text numerically so an application can compare meaning. The Decisions interface is designed for yes/no, multiple-choice and scored questions about text.

For example, a reading extension could compare an article with a project description and highlight relevant passages. It need not produce a long chat response to be useful. Opera's Project Lens demonstration explores that kind of workflow; its example extensions are experiments, not features built into every Opera installation.

How to enable the preview

Opera's setup guide says to update or download Opera One, open Settings, select Opera AI and enable Web AI API. A compatible model then downloads, which can take a few minutes.

Supported platforms are 64-bit Intel or AMD Windows PCs and Apple-silicon Macs. The guide excludes Intel Macs, Linux and Windows on Arm. Performance depends on the device, and local processing consumes its resources.

There is a documentation mismatch worth flagging. The guide is dated October 22 and describes a preview across One, GX and Neon, while the October 27 newsroom announcement says One first and the other two later. The narrower newsroom rollout is the safer availability expectation. The earlier guide also means October 27 should not be read as proof that no preview existed before that date.

What developers still need to handle

Local inference removes a cloud-model round trip from the supported operation. It does not make an entire website private: the site's other network requests, storage and analytics still need their own review. Nor does a model running locally guarantee that its classifications or generated text are correct.

A production feature therefore needs a sensible path when the API is unavailable, the user declines permission or the machine is too slow. A reading tool can fall back to ordinary search and manual filtering rather than quietly uploading the same text elsewhere.

The useful test is whether local AI improves a specific browser task enough to justify the download and device workload. That is a more practical adoption question than whether another chat box can run on a laptop.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and engineering updates into clear stories that developers can understand and share.

Book a 30-minute call.