AI Glasses And AI Backends: Who Actually Runs What

Earlier versions of this article described a specific test methodology — network conditions, query categories, a shielded RF environment — that never took place. No latency benchmark of any AI glasses product has been conducted for this site, and no independently published one was found either. Any millisecond figure you see for this, including earlier versions of this article, was invented.

The Blayzer and Scriber run Meta AI. Meta has not published a parameter count, a tiered on-device/cloud architecture, or a latency figure for it — the specific numbers and architecture described in earlier versions of this article were invented from nothing.

This is worth being direct about: the G2 runs Even Realities' own "Even AI" system, not OpenAI's GPT-4o. Earlier versions of this article attributed it to GPT-4o throughout, including an invented cloud-routing architecture and a latency breakdown for a system the device does not use. The G2 also has no camera, so "visual scene analysis" is not something it does regardless of which AI is involved.

Response speed is a real thing that matters for AI glasses, and it is entirely reasonable to want to know it before buying. Nobody has published it for these products yet. If a real, sourced benchmark becomes available, this page will be updated to reflect it — until then there is nothing accurate to report beyond which AI system each product actually runs.