Claude Code exposes an effort parameter — low, medium, and high — that controls how hard the model works on a given task. Medium is the default, but the documentation doesn't say much about what these levels actually do under the hood. So we ran some tests to find out. Same prompt, 3 terminals, Opus 4.6, everything measured.
The Numbers
| Effort | Cost | Time | Behavior | Models |
|---|---|---|---|---|
| low | $0.14 | 14s | Single model | Opus 4.6 |
| medium | $0.14 | 15s | Single model | Opus 4.6 |
| high | $0.33 | 77s | Multi-model | Opus 4.6 + Haiku |
low and medium are practically identical — same cost, same speed, same output quality for creative tasks.
high is a different story: 2.4× more expensive, 5.4× slower.
But here's what surprised us: high effort didn't just increase the thinking budget. It activated a completely different architecture. Claude Code spun up a second model — Haiku — internally for sub-tasks, running two models in parallel without us asking for it.
Half the cost of the high run ($0.16 of $0.33) was Haiku working behind the scenes.
What about ultrathink?
If you want to temporarily increase reasoning depth for a single prompt without changing your global effort level, you can type ultrathink at the start of your prompt. This extends the reasoning token budget within your current effort level — so even on medium, Claude will think harder on that specific task.
However, if your effort is already set to high, ultrathink won't change the behavior — you're already at maximum reasoning depth.
What this means for how you use effort
That changes how you should think about the effort parameter:
- low / medium — same model, predictable cost, fast
- high — multi-model mode, Claude decides the architecture
- ultrathink — extends reasoning budget within your current effort level, per prompt
So thinking about effort as a simple slider is a misleading conclusion. It's more of a mode switch that defines how Claude solves the problem.
Interested in what we're building? We're continuing to push the boundaries of AI-augmented development at Hedgineer. If you'd like to learn more about our approach or discuss how these patterns could work for your team, reach out to us.