SideNote ProModel selection and reasoning effort controls
The right amount of computation for a task is rarely the maximum available. SideNote Pro puts model choice and reasoning depth where you can change them, instead of deciding for you once and hiding the switch.
Choosing the model
Model is a per-profile field. When your endpoint supports GET /models, the discovered IDs fill a searchable list. When it does not, you type the ID. A new OpenAI profile starts on gpt-6-luna, and the OpenRouter profile Quick Setup creates on openai/gpt-6-luna, the same model under OpenRouter's name. Both are reasonable everyday starting points rather than a recommendation to stop there.
Because profiles are cheap to duplicate, switching model in practice means switching profile from the composer picker. That keeps a fast model and a heavier one both a click away instead of forcing a trip into Settings mid-task.
Reasoning effort
Some models can be told how much internal reasoning to spend before answering. The OpenAI-compatible way to ask for that is a reasoning_effort field on the request. SideNote Pro exposes it as a per-profile setting.
| Setting | What SideNote Pro sends |
|---|---|
| Provider default | No reasoning_effort field at all. The default, and the most compatible option. |
| None | reasoning_effort: none |
| Minimal | reasoning_effort: minimal |
| Low | reasoning_effort: low |
| Medium | reasoning_effort: medium |
| High | reasoning_effort: high |
| Extra high | reasoning_effort: xhigh |
| Maximum | reasoning_effort: max |
Leaving this on Provider default omits the field entirely. That matters, because a provider or model that does not recognise reasoning_effort may ignore the value or reject the request outright. It is not the same as None: None asks a model for no reasoning, while Provider default says nothing and accepts whatever the model does unprompted, which on a reasoning model can be a great deal. Set an explicit level only when you know the model supports it. On the Google Gemini preset, SideNote Pro holds the level to what Google documents for the model: Gemini 2.5 Pro and Gemini 3 models cannot switch reasoning off, so None is not offered for them, and a model Google's pages do not name is sent with Provider default.
Settings › AI providers
Higher is not better, it is different
It is tempting to set everything to maximum and forget it. That is usually the wrong trade. More reasoning means more tokens generated internally before you see a word, which means a slower answer and a larger bill, and on a task with no hidden structure it buys you very little.
| Request | What it actually needs |
|---|---|
| Rewrite this paragraph so it sounds more professional | Fluency. A fast model at low effort. |
| Summarize this email in five bullets | Compression. Speed matters more than depth. |
| Give me ten alternative headings for this section | Variety. Cheap and fast wins. |
| Review this architecture for concurrency risks | Multi-step reasoning. Worth the depth. |
| Debug this failure from these logs and explain the causal chain | Careful inference over many facts. Worth the depth. |
| Compare these strategies and identify hidden assumptions | Structured analysis. Usually worth the depth. |
The pattern is simple: pay for reasoning when the answer depends on connecting several things correctly, not when it depends on saying one thing well. Choosing the right AI model and understanding reasoning effort go through this in more depth.
Measure it on your own work
There is no fixed percentage of latency or money saved by lowering reasoning effort. Results vary with the model, provider, prompt, output length, network and time of day.
SideNote Pro records token counts and local cost estimates per provider and model, so you can compare two configurations on the work you actually do. See usage and cost tracking.