Same rules · four writers
Choose a character. Compare all four models below; scroll inside any phone to move them together.
Raw writer output. Narrated actions and other mistakes remain visible. Bubbles split only at ||| or blank lines; no words are edited. This private trial does not run the app’s separate safety checks. Initials stand in for photos.
What the results show
My recommendation: use GLM for the next focused trial, keeping these exact references and changing only the shared reply-length direction. Its reactions and interest are stronger here, but it still says too much. This is a small sample, not a model verdict.
MiniMax: some lively moments, but Skye and Brooke still narrate actions. Brooke suggests meeting in person; Freja claims a photo. Philip ends with an unfinished question. Not ready.
Gemma: the four personalities differ and usually invite a response. Philip is restrained; Brooke is exuberant. But Brooke and Skye send too many fragments, and Freja sounds more like written prose than texting.
GLM: clear reactions, curiosity and contrasting voices. However, its replies are still long, and all four use “sleep well” in the last turn. The shared ending problem remains.
Qwen: generally fewer messages, particularly for Philip and Freja, but several replies stop without much invitation. It also assumes the user has a cat. Shorter has not meant more engaging.
No model consistently meets all three goals across this sample: distinctive, human texting, and making the person feel wanted. These are unedited writer outputs, not the app’s full safety-controlled delivery.
Cost and writing time
Elapsed time for a complete writer call, not time to first text or full app response time. Each model keeps its saved sampling and thinking settings. These are single runs, not a definitive ranking.
Updated shared rule
Write only what you type and send to this person now. Respond to their message with your own reaction and interest, then leave them room to respond. Use the references for your manner, not their events or facts. Keep the informal grammar, fragments and uneven bursts that fit you, without phonetic accents. Separate sent messages with |||. Actions, gestures, private thoughts and descriptions of typing are not sent messages; leave them out. Stop after your reply and wait for theirs.
Exact prompts and reference examples
One prompt per person, identical across all four models. Every turn receives it plus that model’s own unedited conversation history.