Real web search for every model in Shannon

Shannon can now get a real, cited answer from any model in the lineup, even the ones that have never had search of their own. Ask something that needs a current answer while Auto has you on one of the cheaper models, and instead of guessing or admitting it can’t search, Shannon searches and answers from what’s online right now.

Shannon runs on more than one model on purpose. That’s the Mixture of Models principle: each task goes to whichever one does it best today. That promise only holds if every model in the lineup can do the job, and until now, most of them couldn’t do this one.

Ask a cheap model something that needs today’s answer

Only a few models in Shannon’s lineup could search the web on their own: Sonnet, Opus, and Gemini. Everything else Auto reaches for on lighter, cheaper work — Qwen3-Next, Kimi K2 Thinking, GLM-5, Grok, and gpt-oss-120b among them — had no search at all. A plain question like “what’s the current stable release of Python” landing on one of those models used to get an “I can’t search” refusal, or worse, an answer invented from stale training data. Now it searches and answers from real sources.

Trust it searched, not guessed

Every search shows its work: the queries it ran and the sources it cited, live while it happens and still there if you reload the page later. In one test, a model was asked for Python’s latest release. Its own search query still used a stale version (“Python 3.12 new features”), but it landed on the current answer anyway, because the live results overrode what it remembered. And when a search comes back empty, Shannon says so instead of filling the gap with a guess.

Under the hood

The fallback wraps Gemini’s own web-search grounding as a one-shot call any model can reach for. The agent asks for a search, one Gemini call runs it and returns a grounded answer with citations, and that result comes back as the tool’s output to whichever model is doing your work. A model never gets both paths — one that already searches natively keeps using its own, faster route, and everything else gets the fallback. The citation links point through Google’s own redirect domain rather than the original URL, so the plain source name next to each one (wikipedia.org, python.org) is what tells you where the answer came from.

Try it at shannon.bot, no sign-in needed: ask something that needs a current answer and watch the sources show up.