A good starting point
In Translation settings, look for Recommended for translation. The model menu also highlights an affordable Claude option and higher-cost alternatives. The effort menu marks the recommended level for that exact model. Recommendations and available choices update daily.
You can choose a higher effort. It gives the same translator more thinking time; your translation rules and custom prompt stay the same. It does not turn translation into an automatic reply.
When should I turn effort up?
- Everyday messages: start with the recommended effort, often Fast or Low.
- Ambiguous wording or a complex paragraph: try Medium and compare the translation.
- High, Extra high and Maximum: use them when your comparison shows a useful improvement. More thinking does not guarantee a better translation.
Use your custom prompt for tone, regional wording and formality. Extra effort cannot supply context you have not given the translator. Only the levels supported by your chosen model appear.
What changes the usage?
The request includes ChatterPilot’s translation rules, your custom prompt, the message and any included conversation context. The response and supported reasoning also count. Supported caching is accounted for.
There is no fixed “Medium costs twice as much” rule. A simple message may use little extra reasoning; a harder one may use more. Models also have different prices. Your plan percentage and purchased Lux reflect the measured request cost, so there is no fixed number of translations per Lux.
A measured example
We sent the same 12 synthetic messages at each of four efforts on October 27, 2026, using GPT-5.6 Luna. All 48 replies passed our format checks. We did not grade semantic quality.
| Effort | Median response | Average usage vs Fast |
|---|---|---|
| Fast | 1.4 s | 1.00× |
| Low | 1.5 s | 1.00× |
| Medium | 1.3 s | 1.11× |
| High | 1.5 s | 1.45× |
| Extra high / Maximum | Not measured | |
In this sample, Medium used about 11% more Lux than Fast, and High about 45% more. Medium’s median happened to be faster: short runs vary. These timings cover the provider response, excluding app queues, network to our server and display time. They are examples, not speed or usage guarantees, and do not describe Claude or other models.
Keep it simple
Choose the recommended model and effort first. Adjust your custom prompt for your voice. Raise effort only when comparing results gives you a reason to keep it.
Technical references: OpenAI reasoning documentation and Claude effort documentation. Effort names and supported levels vary by model.
