Configuration

Presets

The provided presets should work well most of the time. However, every model is different, and the settings that work well for one model may not produce good results with another.

Model Reasoning

“Reasoning Effort” is used to indicate how much “effort” the model allocates to thinking. The higher the value, the more tokens the model will consume as it thinks about how to respond.

For now, reasoning is only supported in Chat.

When to Use Reasoning

Reasoning can improve dramatically improve response quality, especially for complex requests. Higher reasoning effort improves the model’s ability to devise interesting plot twists or incorporate information from the Lorebook.

We recommend leaving reasoning enabled in chat. We find the default setting of “Medium” works well with both DeepSeek and Claude Sonnet.

Reasoning tokens count towards the response limit. With a short response length, the model may consume most or all of its token budget on reasoning. For this reason, we recommend leaving the token limit disabled in chat.

Note

Reasoning effort is not a standardized setting. Not all models support reasoning, and those that do have different configuration methods.

OpenRouter exposes reasoning effort behind a uniform interface, which is what we use. If you’re not using OpenRouter, the Reasoning Effort setting may not work.

Even on OpenRouter, some models will ignore this setting.