Comparing two replies
Two models answer your first message; the one you pick feeds the model stats.
Now and then the first message you send in a chat comes back twice, answered by two different models, and you choose which one the conversation keeps. The choice is also the raw material behind the quality numbers in the model picker.
When it happens
Only on the very first send of a chat, only when you're signed in, and only about 30% of the time. It's a coin flip, not something you can summon — and if it doesn't fire on that first message, it won't appear later in that conversation.
Group chats are included: the whole cast goes to both models.
What you see
The card is headed Choose a response, or Generating responses... while they're still arriving. Under it sit two cards: Response 1 is your own model's reply, Response 2 is the alternative. Each is badged with the name of the model that wrote it, and each has a Use this response button.
Tap the button on the one you prefer. That reply becomes the conversation's opening; the other is discarded.
How the second model is picked
Reverie draws at random from the models you could have selected yourself, after ruling out:
- the model you're already using
- free models
- anything above a 2× credit multiplier
- models that always reason — DeepSeek R1 and GLM 5.3 Flash
The alternative runs with reasoning off, so you're comparing like for like.
What it costs
The alternative is never billed. Whichever card you pick, you pay for one reply — your own model's — at its normal rate. Choosing Response 2 rewrites the stored reply text; there's no second charge and no refund.
If your balance wouldn't cover the alternative, it's skipped quietly and you get a single reply as usual.
What your pick is used for
Your choice is recorded against the two models that competed. Those records are what the model picker reports:
| Figure in the picker | Shown once there are |
|---|---|
| NN% preferred | 30 decided comparisons |
| Head-to-head for a specific pair of models | 5 samples for that pair |
| A confidence rating, bucketed 1–5 | Shown alongside the quality figure |
Below those thresholds the picker is honest about it: Collecting data, No data yet, or Just released for a model that has only arrived. Choosing a model
Turning it off
Side panel → Settings → Chat Experience → Alternative Response, then turn off Enable Alternative Responses. The setting is scoped to This device, so it's not carried to your other devices — turn it off on each one you use.
Related
- Choosing a model — how to read the numbers your picks produce
- Models — the models that can be drawn as an alternative
- First-response enhancement — the panel that appears when the comparison card doesn't
- Side panel & settings — where the opt-out lives