Branch the loser, keep the winner: model A/B inside one thread

The single most-used feature in the studio isn’t image generation or video or web search. It’s the small icon under every message that lets you re-run it on a different model. People discover it by accident, and then they stop reading model benchmarks, because they can just check.

The problem with “which model is best”

It’s the wrong question. There is no best model — there’s a best model for this prompt, today. GPT-5.5 writes cleaner first-draft prose. Claude Opus 4.7 holds a long argument together without losing the thread. Gemini 3.1 Pro is unreasonably good at pulling structure out of a messy document. Grok is faster and funnier and wrong more often. The ranking flips depending on what you’re doing in the next five minutes.

The honest way to know is to run the same prompt through two or three of them and read the outputs side by side. The dishonest way is to read a leaderboard from six weeks ago. We built the product around the honest way.

How branching works

Every message has a branch action. Hit it, pick a different model, and the studio re-runs that message with the full conversation context intact — system instructions, prior turns, attached files, everything. You don’t re-type anything. You get a parallel branch you can compare against the original.

Keep the winner and the conversation continues down that path. The losing branches stay in the tree, collapsed, in case you want to come back. Nothing is destroyed; the thread becomes a small experiment log.

A workflow that actually pays off

Here’s the pattern that converts skeptics, lifted from real use:

• Draft on a fast model. Start a piece of writing on a quick, cheap model to get the shape. Cheap credits, fast turnaround.
• Branch the hard turn to a reasoning model. When you hit the part that’s actually difficult — the argument that has to hold, the code that has to be correct — branch just that message to Opus or GPT-5.5. You pay the premium only where it matters.
• Compare, don’t guess. When two models disagree on something factual, that disagreement is information. Branch a third model as a tiebreaker, or send web search after it.

The result is that your average cost per conversation drops, because you’re not paying frontier prices for the easy 80% — and your quality on the hard 20% goes up, because you’re routing it to the model that’s actually good at it.

Why one subscription matters here

Branching only works if the models are all in one place, on one shared credit pool. The moment you’re juggling three separate subscriptions and three separate tabs, the friction kills the habit — you stop comparing and just use whatever tab is already open. That’s how you end up locked into a single provider by inertia rather than by choice.

Inside the studio, switching models is one click and the same credits. The cost of checking is near zero, so people check. Over a few weeks that turns into a real, earned sense of which model to reach for — which is worth more than any benchmark we could publish.

Try it on your next hard prompt

Next time you get an answer that’s almost right, don’t re-prompt the same model and hope. Branch it to a different one. Half the time the second model just gets it — and you’ll have learned something durable about the two of them that no review article would have told you.