Guides
Auto-Switch to Pro
Most conversations stay short and never go anywhere. A few keep going, and those are the ones where output quality decides whether the User stays. Auto-Switch to Pro puts the Smart model on the first kind and the Pro model on the second, with nothing to manage on your side.
How it works
A conversation starts on the Smart tier. It moves to the Pro tier the moment it outgrows its short-term memory window, today the 30 most recent messages, and stays on Pro for the rest of its life.
| Conversation length | Tier used | Long-term Memory |
|---|---|---|
| Up to 30 messages | Smart | Not kept |
| From the 31st message on | Pro | Kept |
The message being processed counts, so the switch happens on the reply to the 31st message. The threshold follows the memory window: if that window moves, the switch point moves with it.
The end user notices nothing
No interruption, no pause, no restart: same Persona, same personality, same history. Tone and continuity carry over unchanged. Which tier produced a reply is a routing decision on our side and is never exposed to the User.
What it costs
Each reply is billed at the tier that produced it: Smart before the switch, Pro after. Smart and Pro usage appear as separate lines on your invoice, so you can see exactly what each part of a conversation cost. See Choose Your Model for both rates, or the rates in your dashboard.
In practice you pay the Smart rate on the bulk of your volume, and the Pro rate only on the conversations that lasted long enough to deserve it.
Turn it on
Point replyParameters.chatModel at the -auto variant of your model version when you enqueue a conversation, for example chat-2.1-auto. The choice is made per request: there is nothing to switch on in the dashboard. See the API Reference.
What Pro adds past the switch
Beyond output quality, the Pro tier is what unlocks Long-term Memory: the facts the AI has learned about a User and a Persona, kept across their conversations. Recording starts at the switch, so a conversation that never reaches the threshold never produces memories. See Conversation Memory for what the short-term window holds, and the Memory Webhook to receive what is learned.