Overview
WebAI is a web technology proposal Opera announced on October 28, 2026. The goal is straightforward: run AI models inside the browser instead of on a cloud server. It is a way of wiring models in rather than a model of its own, and Opera stresses that it is not tied to any specific model and targets open-weight systems generally.
It covers three jobs: generating text, classifying content and understanding relationships between pieces of information. Google's Gemma is the example Opera cites, and on Macs WebAI can also reach Apple's on-device foundation models, which brings the operating system's own layer into scope.
It is an early preview inside Opera One, opt-in for now, with Opera GX and Opera Neon following. Opera frames the economics plainly: developers add AI without standing up infrastructure or paying per request, and users avoid a subscription while their data stays on their device. The caveat is hardware; Opera says real performance depends on the machine.
Key Features
- Model-agnostic wiring: WebAI does not lock in one model. Opera describes it as built for many open-weight systems, so changing models does not mean changing the integration. The browser supplies the runtime.
- Apple on-device models on Mac: Beyond open-weight models, Macs can call Apple's on-device foundation models. Nothing extra to download, which matters for laptop users short on disk or bandwidth.
- Off by default: The preview requires users to enable it. That fits Opera's privacy stance and keeps machines that cannot handle the load from spending compute by default.
- No per-request cost: Opera's own wording is that developers add AI without deploying infrastructure or paying per request. For small tools that pushes marginal cost to zero.
- Data stays on the device: All inference runs locally rather than on corporate servers. For anyone pasting private material into a chat box, that is more direct than any privacy policy.
- Three product lines: Preview first in Opera One, then in Opera GX for gaming and streaming and Opera Neon for agentic browsing.
Use Cases
- In-browser text generation and rewriting without sending drafts to a server.
- Bulk local classification and tagging of a corpus.
- Internal browser tools where inference must not leave the device.
- Mac users calling system on-device models for light tasks, skipping downloads and subscriptions.
- Small web tools embedding AI in the front end with no backend.
- Privacy-sensitive personal use where chat history and notes stay on the machine.
Pros
- Model-agnostic, so one integration works across open-weight models.
- Reaches Apple's on-device models on Mac without an extra download.
- Inference happens locally and data stays on the device.
- No per-request billing for developers to budget around.
- Off by default, leaving compute cost to the user's choice.
- Opera GX and Opera Neon follow, widening where it applies.
Pricing
Ships with the Opera browser at no separate subscription. Local inference incurs no per-call charge; the cost shows up as device compute and power. Currently an early preview in Opera One that users must enable manually.
Summary
The point of this route is converting AI from a per-call expense into a one-time investment. Individuals save a subscription, developers save a backend. The constraints are equally concrete: model size is bounded by device memory, speed by the chip, and Opera itself says performance depends on hardware. Try it by enabling it in Opera One and comparing output quality on text you actually write before relying on it.
Version History
- Opera introduces WebAI: AI models running locally in the browser (2026-10-28): Opera announced WebAI, a way to run open-weight models inside the browser without a cloud service. Opera emphasises model agnosticism, cites Google's Gemma as an example, and notes that Macs can additionally call Apple's on-device foundation models. It is an early preview in Opera One requiring user opt-in, with Opera GX and Opera Neon to follow. Opera says developers add AI without infrastructure or per-request fees and that user data stays on the device, with real performance depending on hardware.