Web search that cites its sources, on every model

A model that won’t tell you where it got something is a model you have to double-check by hand, which means it didn’t actually save you the work. We treat web search as a tool the model calls when it needs fresh or verifiable information — and we make it show its work every time.

Search is a tool, not a mode

You don’t flip a “search mode” switch. You ask your question normally, and the model decides whether it needs to look something up. Ask Claude for last quarter’s numbers and it searches. Ask GPT-5.5 to summarize a debate that’s still unfolding and it searches. Ask either of them something timeless and it just answers from what it knows, no wasted round-trip.

This matters because the model has the context to know when it’s out of date. A separate search box doesn’t. The model that just read your last three messages is in a far better position to decide whether a claim needs a citation than you are before you’ve even seen the answer.

Every claim, every URL

When a model uses search, the answer comes back with inline references — the specific sources behind the specific claims, not a generic “here are some links” block at the bottom. Hover a citation and you see the page it came from. Disagree with the conclusion? Follow the link and check the primary source yourself in two seconds.

The point isn’t to make the model look authoritative. It’s the opposite: it’s to make the model checkable. A cited answer you can verify is worth more than a confident answer you can’t, even when the confident one happens to be right.

Why it works across every model

Because search is a shared tool rather than a feature baked into one provider, you get the same sourced behavior whether you’re talking to GPT, Claude, Gemini, or Grok. The model picks when to call it; the studio runs the search and feeds the results back in; the citations render the same way regardless of who’s answering.

That consistency is the quiet benefit of a multi-provider studio. You’re not learning four different “search experiences” with four different quirks. You learn one, and it follows you across every model you branch to.

A few things it’s especially good for

• Anything time-sensitive. Prices, releases, standings, “what happened this week.” The model reaches for search automatically and you get a dated, sourced answer instead of a stale guess.
• Fact-checking another model. Got an answer that smells off? Ask a second model to verify it with search. The disagreement plus the sources usually settles it faster than you’d settle it yourself.
• Research you’ll have to defend. When you need to paste an answer into a doc someone else will scrutinize, the citations come with it. You’re handing over an argument with its receipts attached.

Included, not metered into oblivion

Web search runs on your subscription. There’s no separate search add-on, no per-query meter that makes you ration questions. The model calls it when it helps, you read the sources, and the credits work the same as everything else in the studio. The goal is for “just check” to be the default instinct — and defaults only stick when they’re free of friction and free of a second invoice.