Can websites use AI without sending page data to the cloud? Opera has shipped an early local preview
Opera 139 now exposes opt-in prompt, classification and embedding interfaces backed by on-device models. The preview narrows cloud disclosure but does not make every model, extension or permission choice private or reliable.
The short version
- What happened
- Opera 139 has launched WebAI as an opt-in early preview: compatible sites and extensions can call local prompt, classification and semantic-embedding interfaces.
- Who it affects
- A person could use a page-aware AI tool without automatically sending the relevant text to a cloud model provider. They still need understandable controls over which sites can call the model and evidence that extensions cannot misuse the local capability.
- What remains unproven
- Company announcement
Keep in mind
Company announcement
Early preview
No independent benchmark
No published energy or battery measurements
Limited operating-system and processor support
Model and hardware performance varies
Local processing does not validate calling sites or extensions
Source analysis · 2 linked sources. Primary analysed source date: 27 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 3 min The evidence, methods, implications and limitations
The browser becomes the model host
Opera 139 now lets a compatible website or extension ask the browser to generate text, classify information or compare text semantically using an on-device model. The user must first enable WebAI in settings and download a compatible model. Opera says permissions are granted per site, and developers do not need their own cloud AI account, server or per-request API bill.
That design can keep the page material handled by the local model on the user's computer instead of sending it to an external model provider. It also shifts compute, storage and electricity use onto the device. Response speed and capability therefore depend on the selected model and hardware, not just the website calling the interface.[1][2]
Local does not automatically mean safe or dependable
Opera calls WebAI a proposed standard and an early preview. Its examples are experimental extensions, not built-in production features. The launch supplies no shared test of hallucination, malicious prompts, permission confusion, model updates, battery use or performance across supported computers. Sites still decide what they ask the model to process, while users must understand and manage access.
The first release supports 64-bit Windows PCs with Intel or AMD processors and Macs with Apple silicon; Linux, Windows on Arm and Intel Macs are excluded. Independent browser-security review, clear permission telemetry and reproducible model-and-device benchmarks would show whether the local approach delivers its privacy and usability promise at scale.[1][2]
The distinction that matters
Where the data travels
The preview changes who supplies the model and where inference runs.
Enable
User opts in and downloads a model
No model download by default
Permit
A site or extension requests WebAI access
Site-specific browser permission
Run
The browser processes supported tasks locally
No required cloud-model round trip
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 2 min 31 sec · Background explainer
What is artificial intelligence? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Published 2018-11-28. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current AI technologyPhoto: ThisisEngineering / Unsplash · Licence
Continue the story
Read next
Continue exploring privacy: compare the finding and its evidence limits.
Can AI agents share lessons without sharing private traces?
FedSkillGit replaced whole-document sharing with filtered rule-level patches and beat independent learning in nine benchmark combinations. Its privacy evidence measured an AI audit, not resistance to extraction attacks.
Read the next reportExplore the full source and evidence record
What this means for people
- A person could use a page-aware AI tool without automatically sending the relevant text to a cloud model provider. They still need understandable controls over which sites can call the model and evidence that extensions cannot misuse the local capability.
Global context
The preview is distributed through a global browser but initially depends on a narrow set of desktop platforms. Whether other browsers converge on compatible interfaces will determine if this becomes a portable web capability or an Opera-specific experiment.
What the evidence does not yet show
- Company announcement
- Early preview
- No independent benchmark
- No published energy or battery measurements
- Limited operating-system and processor support
- Model and hardware performance varies
- Local processing does not validate calling sites or extensions
What to watch next
- Independent security review
- Permission design and abuse reporting
- Cross-browser standardisation
- Model provenance and update controls
- Latency and energy benchmarks by device
- Linux, Arm and older-device support
Living evidence record
Impact record IAI-0B2515Y
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
28 October 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 28 October 2026
Research question, topics and regions
Topics: Browsers · On-device AI · Privacy · Web standards · Open-weight models · Developer APIs
Regions: Europe · Norway · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.