Skip to main content

​
Overview

Whim gives you two controls to tune the speed vs quality tradeoff: fast mode (available for Claude and Codex) and reasoning effort (Codex only). These let you optimize for quick iteration when speed matters, or deep reasoning when accuracy matters.

​
Fast Mode

Fast mode tells the AI provider to prioritize response speed over depth of reasoning. It’s useful for straightforward tasks where you want rapid output — simple edits, boilerplate, quick fixes, and iterative changes.

​
How It Works

ProviderEffectSupported Models
ClaudePasses --effort low to the CLI, reducing reasoning depthAll Claude models (Opus 4.6, Sonnet 4.5, Sonnet 4.5 1M, Haiku 4.5)
CodexSets service_tier=fast in the Codex configGPT 5.4 only
CCR + OpenRouterNot available—
Fast mode for Claude uses the same model — it does not switch to a different, smaller model. It reduces reasoning effort to produce faster output.

​
When to Use Fast Mode

  • The task is straightforward (renaming, formatting, simple bug fixes)
  • You’re iterating quickly and want rapid feedback
  • You’re running multiple tasks in parallel and want to conserve time
  • The task doesn’t require deep architectural reasoning
  • The task involves complex logic or multi-step reasoning
  • You need the agent to carefully analyze trade-offs
  • The task requires understanding large amounts of context
  • Accuracy is more important than speed

​
Enabling Fast Mode

You can enable fast mode at three levels:
  1. Per task — Toggle fast mode when creating a task. This overrides all defaults for that task only.
  2. User default — Go to Settings > My Defaults and enable fast mode under the relevant provider section. Applies to all your new tasks.
  3. Workspace default — Go to Settings > Workspace > Defaults and enable fast mode. Applies to all new tasks for all workspace members (unless overridden by a user default).
Codex has fast mode enabled by default for GPT 5.4. Claude has fast mode disabled by default.

​
Reasoning Effort (Codex Only)

Reasoning effort controls how much compute the Codex model spends thinking through a problem before responding. Higher effort means more thorough analysis but slower output.

​
Levels

LevelBest forSpeed
LowSimple, well-defined tasks. Minimal reasoning needed.Fastest
MediumModerate tasks. Some reasoning helpful but not critical.Fast
HighComplex tasks that benefit from deeper analysis.Moderate
xhighDifficult problems requiring extended reasoning — complex debugging, architecture, multi-file refactors.Slowest
The default reasoning effort for Codex is xhigh. This gives the best results for most coding tasks. Lower it when you want faster output on simpler work.

​
Setting Reasoning Effort

Like fast mode, reasoning effort can be set at three levels:
  1. Per task — Choose a reasoning effort level when creating a task
  2. User default — Set in Settings > My Defaults > Codex
  3. Workspace default — Set in Settings > Workspace > Defaults
User defaults override workspace defaults. Per-task settings override both.

​
Speed vs Quality Reference

Here’s how the different controls compare across providers:
ConfigurationSpeedQualityCU Impact
Claude Opus 4.6 + fast mode offSlowerHighestMore container time per task
Claude Opus 4.6 + fast mode onFasterHighLess container time
Claude Sonnet 4.5 + fast mode offFastHighModerate container time
Claude Sonnet 4.5 + fast mode onFastest (Claude)GoodLeast container time
Claude Haiku 4.5Very fastGoodLow container time
Codex GPT 5.4 + xhigh reasoningSlowerHighest (GPT)More container time
Codex GPT 5.4 + low reasoningFastest (Codex)LowerLeast container time

​
CU Impact

Fast mode and lower reasoning effort reduce the time the agent spends on each task, which directly reduces container runtime CU costs. For subscription providers (Claude, Codex), this is the only CU cost. For CCR + OpenRouter, it also reduces token usage since the model generates less reasoning output. The tradeoff is straightforward: faster responses use fewer CU but may produce lower-quality results on complex tasks.

​
Next Steps

Model Selection

Choose the right model for your task.

Compute Units

Understand how CU costs work.