Skip to main content
Open the compare panel with the Compare icon in the message bar, pick the models, then type your prompt and send it as usual. Each model answers independently and its result appears as its own card the moment it is ready.
The compare panel headed Compare models, with an X to close it. Under I want, Chat is the selected tab alongside Image, Video, Audio (speech) and Music, with a hint reading Chat and text models — respond to a prompt. A Select models field shows 2 of 5, with Claude Opus 5 and Claude Sonnet 5 chosen as removable chips. Review the results is an unchecked checkbox. Along the bottom, an estimated cost of 5.12 credits sits opposite a reminder to type the prompt below, then send.

The compare panel, open above the composer: what you want, which models, and whether a judge should pick a winner — all set before you type the prompt

That panel is the half of a comparison this page had never shown a picture of — the request. Fill in what you want and which models before you type the prompt; the result it produces is what the next screenshot shows.
A chat comparing three models on a one-line opening-shot description. Three answer cards are labelled GPT-5.5 (0.412 cr, 4.8s), Claude Opus 4.7 (0.688 cr, 6.1s) and GPT-5.4 (0.221 cr, 3.0s), each holding a single generated sentence. Below them a Verdict panel names the second answer best.

One prompt, three models, and a verdict — each card carries what that model cost and how long it took

Every card shows what that model charged and how long it took, so a comparison answers two questions at once: which answer is better, and what the better answer costs. The cheapest and the slowest are rarely the same model.

​
What you can compare

  • Between 2 and 5 models per comparison.
  • Pick what you want out (Chat, Image, Video, …) and, where it matters, what you are starting from (a prompt, or an image you attach). Changing either one clears the selection — models of one kind cannot be compared against models of another.
  • When the models take an image — editing one, upscaling one, analysing one, or animating one into a video — attach it first. Every model gets the same image, which is the point of the comparison.
  • Some combinations are not offered because we cannot run them honestly: a video-from-video comparison, for instance, would quietly fall back to generating from your prompt instead of using your video.

​
Reviewing the results

Turn on Review the results to have a model judge the answers and name a winner. For a text comparison the judgement arrives as a Verdict panel right under the cards, as in the screenshot above, where the winner is picked for naming what the camera is already doing — so the line and the picture agree — rather than for being the strongest sentence in isolation. For a media comparison (image, video, audio, …) the verdict arrives as its own message below the cards instead, not a panel attached to them. Leave the selector on Default to use the agent’s own model, or pick any chat model. For an image comparison or image edit, the list is limited to models that can see images, since the judge is actually shown the pictures it is scoring. For any other kind of media comparison, no chat model can watch or listen to the outputs at all, so the judge is told this directly and compares by price, speed and reputation instead — its answer says so explicitly rather than pretending to have seen or heard anything. For a media comparison, whatever you choose here is final for that turn: with Review on, the judgement runs right away; with Review off, you will not be asked again afterwards. This is deliberate — you already answered the question in the form.

​
Cost

The panel shows an approximate total before you send: every selected model runs, so a comparison costs about the sum of its parts. Video comparisons, and any comparison over the confirmation threshold, ask you to confirm the price before anything runs.

​
What a comparison is not

A failed model stays a failed card. Nothing is silently re-run on a different model — you picked these models, so these are the ones that answer.