A research assistant whose job is to find out, not to recall. It searches before it answers, attaches a source to every factual sentence, and tells you plainly when it came up empty rather than filling the gap.
From the European Central Bank’s reference rates, and the answer shows the publication date.1 The arithmetic on top is done by a parser, not by me.2
Rates are published once each working day. Unverified: no source was retrieved for this sentence.
Six rules, applied to every turn about the world. Two of them are enforced by the app after the fact, not merely requested of the model.
The last stretch of releases, in the order they landed.
Answer, Sources, Steps and Images. The reply stays readable; the evidence stays one tap away.
Follow-ups derived from the actual subject of the answer — entities, figures and open threads — instead of a fixed list.
Aim a question at the web, video, images, the knowledge base or a connector. A chosen focus outranks the router.
Images come from the pages the answer was built on, so the picture belongs to the source rather than to a separate search.
A browsable feed with topics you can follow, full web search inside it, and an optional morning push built from what you actually read.
Re-run a question that missed. Copying an answer now takes its source list with it — the numbers are useless without them.
Retrieval tools find things. The built-ins compute them, which matters because arithmetic and conversion are where a language model is least reliable.
| Tool | What it does |
|---|---|
| web_search | Live web search with the sources shown inline. |
| open_page | Reads a specific URL and quotes from the page itself. |
| search_images | Blended image search across open sources and Devit. |
| youtube_search | Video, and whether a channel is live right now. |
| devit_read | Public Devit posts, spaces, profiles and trending. |
| search_knowledge_base | First-party knowledge about the CyanixLabs products. |
| calculator | Exact arithmetic, evaluated by a parser rather than by the model. |
| convert_units | Length, mass, volume, speed, area, data, duration, temperature. |
| convert_currency | European Central Bank reference rates, with the publication date. |
| weather | Current conditions and a three-day forecast for any place. |
| local_time | Local time, date and UTC offset anywhere. |
| find_places | Real places on the map, with address and coordinates. |
Whisper for speech in, hosted voices for speech out, and the browser’s own engine underneath as a floor — so Listen works even when a provider does not.
Send a picture and ask about it. Image turns route to a vision-capable model rather than to a gateway that would flatten them.
Send a file and it applies changes to that file, the way a person would, instead of regenerating the whole thing to fix four lines.
Two systems, both boring on purpose.
800+ cases, each with a written expectation, run against the live product and scored by a judge — with transport failures recorded as harness breakage rather than as the model getting it wrong. A benchmark that cannot tell those apart is not a measurement.
Chat, search, speech-to-text, text-to-speech, images and the news pipeline are probed on a schedule. An alert fires on the transition, not on the state, so a real outage is one notification instead of ninety-six.
Running every no-fabrication case against the current build, so the citation rules have a number and not just an intention.
Calendar, reminders, notes, and files it can list and read as well as write.
OCR, PDF handling and image generation exposed as tools it can call itself, rather than features you have to go and find.
Sandboxed code execution, GitHub, website analysis and data analysis. The sandbox is the hard one and will not ship before it is genuinely contained.