Effort levels

Effort levels

Effort levels are one-click presets that tune how hard the model thinks before replying. They map to reasoning-effort, max-tokens, and sub-agent budgets in the background, so you don’t have to configure each provider individually.

The selector lives next to the Send button. It applies to the next message only — pick a level, send, and the composer returns to your default.

The three levels

Level	When to use	What it does under the hood
Quick	Simple edits, one-liners, “what does this do?”	Minimal reasoning effort, tight max-token budget, no sub-agents. Fastest and cheapest.
Balanced	Normal development work — refactors, small features, code review.	Default reasoning effort, standard budget, sub-agents allowed but not encouraged.
Thorough	Multi-file features, debugging hard issues, architectural design.	High reasoning effort, generous budget, autopilot encouraged to spawn specialists.

How effort maps per provider

• Claude (Anthropic API & OpenRouter) — adjusts thinking.budget_tokens and max_tokens.
• GPT-5 / Codex — sets reasoning.effort (low, medium, high).
• Gemini — adjusts thinking budget on 2.5-series models.
• Ollama / local models — adjusts context size and sampling where supported.
• Copilot — maps to the underlying upstream model’s effort controls when the selected Copilot model exposes them.

Note

Providers that don’t expose an effort knob (e.g. older chat-completions models) ignore the level but still receive the corresponding max-token budget.

Setting a default

Your default effort level is per-workspace and lives in workspace settings. The in-chat selector overrides the default for one message at a time.

Tip

Thorough costs noticeably more than Quick — often 3–10x. Keep an eye on the cost bar when you use it.