Reasoning Settings Explained:
DeepSeek & OpenAI in OpenClaw

What Is "Reasoning" in an AI Model?

When you send a message to an AI model, the model doesn't just spit out an answer. With reasoning (also called "thinking"), the model works through the problem internally before giving you a final response. It's like the difference between someone blurting out an answer versus someone saying "let me think about this for a moment."

In OpenClaw, this internal work is normally hidden. You see only the final answer. But you can control both how much thinking the model does and whether you see it using two separate settings:

The Thinking Levels

OpenClaw defines a standard set of thinking levels. What each level means depends on the provider, but the names are consistent across all supported models:

How to Use These Settings

Thinking directives work in two ways:

/think high

/think xhigh Can you analyze this contract and find any red flags?

/think

To clear a session override and go back to the configured default:

/think default

Aliases: inherit, clear, reset, unpin all work the same way.

DeepSeek V4: What's Really Happening

DeepSeek V4 models (deepseek/deepseek-v4-flash and deepseek/deepseek-v4-pro) handle reasoning differently than most providers. Here's what you need to know:

Thinking and Tool Calls

DeepSeek V4 has a stricter contract around thinking with tools. When the model thinks during a turn and then uses a tool, DeepSeek expects the assistant's internal reasoning (reasoning_content) to be included in the next request. OpenClaw handles this automatically — you don't need to do anything special for multi-turn tool use with thinking enabled.

When Thinking Is Off

When you set /think off with DeepSeek V4, OpenClaw sends thinking: { type: "disabled" } and strips any replayed reasoning_content from the conversation history. This keeps the session cleanly on the non-thinking path.

Which Model to Use

OpenAI (GPT-5.5): What's Really Happening

OpenAI's GPT-5 family models support reasoning through the Responses API. Unlike DeepSeek, OpenAI offers a wider range of effort levels:

GPT-5.5 Thinking Levels

Seeing the Thinking: /reasoning

Separate from how hard the model thinks, you can control whether you see that thinking:

Which Setting Should You Use?

Here's a quick decision guide. Click your scenario:

How OpenClaw Decides Your Thinking Level

When you send a message, OpenClaw resolves your thinking level in this order:

Configuring Defaults (openclaw.json)

To set a permanent thinking default rather than using session overrides:

{
  "agents": {
    "defaults": {
      "thinkingDefault": "medium",
      "reasoningDefault": "off"
    }
  }
}

This sets /think medium as the starting level for all agents, with reasoning visibility off.

Bonus: Fast Mode (/fast)

Separate from thinking depth, /fast controls processing priority — not reasoning effort:

Common Questions