CX AI models: Auto, Smart and Fast
The free built-in CX AI, its three models, daily quota, Power mode and automatic fallback.
CX AI is CX Agent's own AI service. It is free, needs no API key, and is selected by default on new installs. Requests go to the CX backend, which picks the best available free model for each message and fails over between providers on its own.
The three models#
| Model | Id | Best for | Abilities |
|---|---|---|---|
| CX Auto | cx-auto |
Everyday use. The backend chooses per message. | Text + images, tools |
| CX Smart | cx-smart |
Harder reasoning, planning, coding. | Text + images, tools, reasoning |
| CX Fast | cx-fast |
Quick short answers. | Text, tools |
Pick a model from the model picker in the composer. Your global default is set in Settings → AI models → Default models.
Daily quota#
The CX AI card (in the model picker and Providers) shows "x% used · tokens today / quota" and Resets in …. Quotas depend on your tier (guest, verified, Pro). Sign in to double your daily quota.
When the limit is reached the composer shows Limit reached with the reset time in IST and two ways forward: Choose your own model or Use your own API key. Sometimes only fast models remain until the reset ("Fast models only until the reset").
Power mode#
Power mode is always on (there is no switch for it in this version). Every message tells the CX backend whether it looks like a coding task or a hard reasoning task, so the strongest free models are tried first. Longer messages (over 1,500 characters) and words like prove, solve, plan, compare, design, samjhao count as hard tasks and also get more reasoning budget.
Automatic fallback#
CX never wants a task to die on one model:
- If a reply fails before any text arrives, or no first word arrives in time (10 s for your own providers, 65 s for CX AI which fails over internally), CX retries once on a fallback model.
- The fallback is your global default chat model if it differs, otherwise another CX AI model. For CX AI itself the same model is retried and the backend routes to a different upstream provider.
- It never retries after text has started, so you never get duplicated text.
- A small "Continued with …" note shows when a switch happened.
Context-length errors are not retried (a different model would fail the same way). Use Compress history instead.
Lean mode for short messages#
Short plain messages (no attachments) on CX Auto / CX Fast get a lighter system prompt so answers come back faster.