OpenClaw's /think command controls how much reasoning effort the model applies before generating a response. It's one of those features that's easy to ignore until you either want better answers or you start seeing errors about it.

New in 2026.4.5, thinking levels are also exposed in the Control UI as a setting you can toggle per-session without typing a command. But the command interface still works and is more precise.

The Six Thinking Levels

Level What It Does Token Cost Speed
/think off No extended reasoning. Model answers directly from training. Fastest, cheapest. Some models require a non-off level. Fastest
/think minimal Very lightweight reasoning pass. Catches obvious errors without significant overhead. The safe default for most setups. Fast
/think low Light reasoning. Good for factual questions, straightforward tasks. Noticeable improvement over off for ambiguous requests. Moderate
/think medium Balanced. The most common practical setting. Better multi-step reasoning without excessive cost. Moderate
/think high Extended reasoning before responding. Noticeably better on complex tasks, planning, and ambiguous instructions. Higher cost and latency. Slow
/think max Maximum reasoning budget. Best quality, highest cost. Use for critical one-off tasks where quality matters more than speed or cost. Slowest

When to Use Each Level

/think off

Use for:

  • Simple factual lookups
  • Formatting existing content
  • Heartbeat / background tasks
  • Models that don't support reasoning
/think minimal

Use for:

  • General daily assistant work
  • Default for most conversations
  • Models that require non-off (safe fallback)
/think medium

Use for:

  • Writing tasks needing structure
  • Multi-step research queries
  • Code generation and review
/think high / max

Use for:

  • Complex planning and strategy
  • Debugging tricky code issues
  • High-stakes one-off decisions
  • Tasks where you'd otherwise iterate 3–5 times

The "Reasoning is Required" Error — and How to Fix It

After the 2026.3.28 update, some users started seeing this error on every call:

Reasoning is required for this model endpoint. Use /think minimal (or any non-off level) and try again.

This isn't a bug in OpenClaw — it's a model-side requirement. Some model endpoints (particularly newer reasoning-first models from Anthropic and certain OpenAI-compatible providers) require extended thinking to be enabled. They won't process requests with /think off.

Fix 1 — Set a default thinking level in your config:

"agents": {
  "defaults": {
    "thinking": "minimal"
  }
}

This sets minimal as the default for all sessions, which satisfies models that require a non-off level without significantly increasing cost.

Fix 2 — Set it per-model:

"models": {
  "providers": {
    "anthropic": {
      "models": [
        {
          "id": "claude-opus-4-6",
          "thinking": "minimal"
        }
      ]
    }
  }
}

Fix 3 — Set it in the current session:

/think minimal

This only applies to the current session and resets on restart. Use the config approach for a permanent fix.

Why this error appeared after 3.28: That update changed the default thinking level from minimal to off for performance reasons. Models that previously worked fine now fail because their endpoint requires a non-off level. The config fix above restores the previous behavior.

Setting a Per-Session Default in SOUL.md

If you want different thinking levels for different types of work without typing /think every time, you can instruct your agent in SOUL.md:

# In SOUL.md
## Reasoning
- Use /think minimal for routine tasks and daily questions
- Use /think high for planning, strategy, or anything I explicitly say is important
- Use /think max only when I say "think hard about this" or similar

This gives the model context to self-select the appropriate level based on task type, rather than applying one level uniformly to everything.

Thinking Levels and API Cost

Extended thinking uses additional tokens — both input (the reasoning prompt overhead) and output (the reasoning trace, which is billed even if it's not shown to you). Rough multipliers:

For a setup spending ~$3/day at off, switching to minimal adds maybe 20–30 cents. Switching to high for everything could triple your bill. Use the higher levels intentionally.

Practical recommendation: Set minimal as your config default to avoid the "reasoning required" errors. Use /think high or /think max explicitly for tasks that need it. Never set max as your default — it will meaningfully increase costs for no benefit on simple tasks.