Opera WebAI: local AI in the browser, free inference per call

Opera now ships local AI models inside its browser, and it lets websites and extensions call them. The feature is called Opera WebAI. A site can ask the visitor's own machine to generate text, classify content or compare meaning, with no API key, no per-request bill and no data leaving the device. It is an early preview, it is opt-in, and it runs in one browser family. That makes it a nice extra, not a foundation. Here is how we would use it.

What actually happened

Opera announced the launch on October 27 in a press release, with TechCrunch covering it. The details:

• The standard. Opera calls WebAI "a proposed web standard" backed by companies including Google and Microsoft. Opera's version is model-agnostic.
• The models. Google's Gemma at launch. On Macs, Apple's on-device foundation model too.
• The APIs. Opera's developer post names three: the Prompt API, the Decision API and the Semantic Embeddings API. Full docs live at next.opera.com/web-ai.
• Where it runs. Early preview in Opera One on 64-bit Windows (Intel or AMD) and Apple silicon Macs. Not on Intel Macs, Linux or Windows on Arm. Opera GX and Opera Neon are next.
• Permissions. Nothing downloads until the user turns on "Enable Web AI API" in opera://settings/ai. Then site-by-site permissions decide which websites can use it.

Opera's pitch to developers, per TechCrunch: add AI features "without running AI infrastructure or paying for every request." Opera also warns that performance depends on the user's hardware. It gives no model sizes.

Why local AI in the browser matters for your business

Small AI features are where cloud bills creep. Tag a support message. Sort a product review as positive or negative. Suggest a related article. Each call costs a fraction of a cent, and then you have 40,000 of them. A local model on the visitor's laptop does that work for free, and the text never hits your server, which helps when the input is something a customer typed.

But only a slice of your visitors run Opera, fewer will opt in, and fewer still have the hardware. So treat WebAI as progressive enhancement:

1. Feature-detect. Check for the API. If it is missing or permission is denied, fall back.
2. Keep the server path. The same task goes to a small hosted model. Same input, same output schema.
3. Use it for low-stakes work. Classification, drafts, search ranking. Never pricing, refunds or anything legal.
4. Validate the output. A local model on an old laptop is a different model from your eval set. Check the shape before you use it.

Done this way, every opted-in visitor is a call you did not pay for, and nobody else notices a difference.

Sources: Opera press release, Opera developer blog, TechCrunch.