Your agent spends tokens thinking before it answers. Extended thinking mode produces better reasoning—but it also produces 2–3× more output tokens per response. For routine tasks, that's expensive overthinking.
The Problem: Thinking Mode Is Always On
OpenClaw supports "thinking" or "extended reasoning" modes that let the AI work through complex problems step-by-step before responding. This is invaluable for coding, debugging, and multi-step analysis.
But when thinkingDefault is set to anything above "minimal," every request gets extended reasoning—including:
- "What time is it in Tokyo?"
- "Summarize this 3-line email"
- "Acknowledge this notification"
These tasks don't benefit from deep reasoning. They just generate extra output tokens that you pay for.
How ClawBridge Detects This (Diagnostic A05)
The Cost Control Center checks your thinkingDefault setting and calculates:
- Current thinking token ratio: What percentage of your output tokens are "thinking" tokens vs. final response tokens?
- Potential savings: If thinking were set to "minimal," how much would you save on output tokens?
The diagnostic triggers when:
thinkingDefaultis not set to "minimal"- Output token costs exceed a configurable threshold
One-Tap Fix
Tap Apply to set thinkingDefault: "minimal". This doesn't disable thinking—it simply makes the AI use the shortest reasoning path for routine tasks. On complex tasks, it still reasons thoroughly.
Trade-offs
- Complex coding tasks: If your agent primarily writes and debugs code, extended thinking produces measurably better results. Consider keeping it enabled if code quality is your top priority.
- Not binary: "Minimal" doesn't mean "no thinking"—it means "think only as much as needed." The AI still reasons; it just doesn't show its work as extensively.
- Model-dependent: Some models (like Opus) benefit more from extended thinking than others. On Haiku, the difference is often negligible.
Real Numbers
Average agent producing 50K output tokens/day with thinking mode on:
- With extended thinking: ~60% of tokens are thinking → 30K thinking tokens wasted
- At Sonnet output pricing ($15/1M output tokens): 30K × $15/1M × 30 days = $13.50/mo savings
- At Opus output pricing ($75/1M output tokens): 30K × $75/1M × 30 days = $67.50/mo savings
FAQ
Q: Will my agent produce lower quality responses? A: For routine tasks, no. For complex multi-step reasoning, you may notice slightly less detailed explanations. The final answer quality is usually the same.
Q: Can I enable thinking for specific tasks only? A: Not through a global setting, but you can instruct your agent in SOUL.md to "think carefully" only when explicitly asked.
ClawBridge is free and open source (MIT License) — install it in seconds, own it forever. Get ClawBridge Free →