Playground · Conversation lab
Conversation lab
Test how different models handle the same conversation, the way you'd test a dialogue design. Replies appear side by side so you can compare them.
Every bot conversation comes down to small calls: what to say first, whether to ask or just help, when to laugh it off and when to stay quiet. I design those conversations, and my real process is far bigger than one page, so this is a preview. Pick a scenario a real person might say, run it across a few models, read the replies side by side, set one as your baseline and see where they differ. Then decide what you'd do.
Try one of the scenarios below, or write your own.
How to read this: same scenario, several models. Read each reply, then compare speed, size and cost in the trade-offs below.
This shared comparison has expired.
Models
Up to 3 at once, or 4 if you allow it. 3 picked.
Scenarios
Start with the angry-customer scenario, or pick your own.
Load a sample conversation
What would a user say?
0 / 1,500 characters. Sample conversations are invented, no real people.
That looked like an instruction aimed at the models, so it was declined. Try writing it as something a user would say.
Back tomorrow. Your past runs are still here.
The models are busy right now. Try again in a minute, or open a past run or a shared link.
Pick a scenario or write your own, then run it to see how each model replies.
Responses
Trade-offs
Benchmarks
A plain checklist for this sample conversation, not a score.
Cost: estimates from listed prices; free-tier limits apply.
Past runs
Saved on this device only. Keeping the latest 10.
Removes past runs saved in this browser. The run on screen stays.
Past runs cleared.
A preview, not the full process. The real thing has many more layers.
How this lab stays within limits
- Runs on free tiers, so each visitor gets 30 runs a day.
- Identical runs reuse a saved result for a day.
- Nothing about you is stored beyond this device. Shared links expire after 30 days.
- Replies are shown as plain text, never run as code.