What happened
On October 27, 2026, Opera announced WebAI, an early-preview implementation of local AI for websites and browser extensions. Opera describes WebAI as a proposed web standard backed by companies including Google and Microsoft. Its premise is straightforward: a web developer should be able to call AI capabilities from a page or extension while inference runs in the user’s browser, on that user’s computer.
Opera says its implementation can use open-weight models running on Opera’s inference engine as well as compatible models already present on a device. The company frames this as an alternative to installing a separate AI application, configuring a local server, obtaining an API key, or sending each request to a cloud provider. The stated benefits are that processed page data can remain on-device, developers do not face per-request cloud charges, and applications are not tied to one model provider.
The preview exposes three developer-facing interfaces. Prompt API is intended for text generation. Decision API is intended for classification. Semantic Embeddings API is intended for finding relationships among pieces of content. Opera’s examples include an extension that ranks job postings, prioritizes articles, highlights material matching a person’s interests, or makes saved content searchable. The company also built six experimental extensions, including Project Lens, Got It?, Questionable Intelligence, Semantic Art, RecordIn Alchemy, and C·I·A·O. Opera presents them as demonstrations rather than browser features.
WebAI is opt-in. Nothing downloads until a user turns it on in Settings, and the feature can be disabled later. Opera’s test instructions say to update or download Opera One, enable Web AI API under Settings → Opera AI, wait for a compatible local model to download, then install an example extension. The preview is available on 64-bit Windows PCs with Intel or AMD processors and Macs with Apple silicon. It is unavailable on Intel-based Macs, Linux, and Windows on Arm. Because inference uses local processing power, response speed depends on the device; Opera says performance work is ongoing.
What the archive adds
This is a new event, not a restatement of an earlier Field Note. Earlier Opera browser work connected external AI systems to browser context through an MCP-based Browser Connector; WebAI changes the direction of travel by letting sites and extensions invoke local browser AI. The archive’s Decisions API coverage concerns a different product and is not evidence about Opera’s implementation.
Trend signal
WebAI supports the pattern that agent-ready software exposes capabilities. Opera is not merely adding an assistant surface: it is exposing bounded functions—generation, classification, and embeddings—to applications running in the browser. The opt-in setting and local model download are also concrete product boundaries, although the announcement does not establish a complete permission model for every possible extension or site use.
Why now
Opera says it first gave developers access to local models in an experimental browser build in 2024. It argues that progress in open-weight models and consumer hardware has now made local execution more practical, while developers increasingly want to avoid routing every AI request to the cloud, paying per token, and selecting a single model vendor.
Connections to other events
Opera previously introduced Browser Connector, which uses MCP to give ChatGPT or Claude access to browser context such as tabs, page content, screenshots, and navigation. That approach brings outside models to browser activity. WebAI instead offers applications within the browser local inference primitives. Opera also launched the Neon AI browser waitlist earlier; Neon is one of the browsers named for this WebAI preview.
Why it matters
For teams building browser extensions or AI-enhanced web experiences, WebAI proposes a different deployment shape. A capability such as local classification of page content or semantic search over saved material may not require a hosted inference service, a user API key, or a metered model call. That can simplify privacy-sensitive features and remove an operational dependency for certain workloads.
The trade-off is that the runtime becomes heterogeneous. Users must explicitly enable the feature and download a model; supported operating systems and processors are limited; and performance follows the user’s hardware rather than a controlled server fleet. Builders should therefore treat WebAI as an optional local execution path, with graceful behavior when it is disabled, unavailable, or too slow for an interaction.
Limits and uncertainty
This is an Opera announcement about an early preview, not proof of broad browser interoperability or production readiness. The source calls WebAI a proposed standard and says Google and Microsoft back it, but it does not specify compatible implementations, conformance requirements, model sizes, quality levels, memory requirements, or a timeline for standardization. It also does not establish what controls a site or extension receives over model choice, permissions, input scope, or output safety. The announced availability is limited to specified Windows PCs and Apple-silicon Macs, and Opera explicitly says local performance is still improving.
Skill to build now
Pick one extension or web feature whose core operation is bounded: classify the current page, generate a short transformation, or retrieve related saved items by semantic similarity. Define a non-AI fallback first, then test the same workflow with WebAI enabled on a supported device. Measure the interaction latency and make the enabled/disabled state visible in the product. This turns the preview into a concrete architecture experiment rather than an assumption that every browser user has local inference available.
Reader takeaway
Opera is making local browser inference callable by websites and extensions through three focused APIs. That can eliminate cloud calls for some tasks, but it is currently an opt-in, hardware-dependent preview rather than a portable assumption for the web. Build for it as a progressive capability, not as the only path.
Open question
Will WebAI acquire interoperable implementations and sufficiently explicit permission boundaries for developers to rely on it beyond Opera’s preview browsers?