Problem

Currently, OpenClaw uses OpenRouter to route to Anthropic models (Claude), but doesn't leverage Anthropic's prompt caching feature. This results in significantly higher costs - potentially 90% more than necessary.

Solution

Add support for Anthropic's cache_control breakpoints in messages sent through OpenRouter. This would allow OpenClaw to cache large static context blocks like:

• System prompts
• Workspace bootstrap files (SOUL.md, USER.md, AGENTS.md, etc.)
• Memory search results
• Large skill documentation

Implementation Details

Per OpenRouter's prompt caching docs, Anthropic caching is supported but requires explicit cache_control breakpoints in message content.

Example:

{
  "messages": [
    {
      "role": "system",
      "content": [
        {
          "type": "text",
          "text": "You are an assistant..."
        },
        {
          "type": "text",
          "text": "<large workspace context>",
          "cache_control": {
            "type": "ephemeral",
            "ttl": "1h"
          }
        }
      ]
    }
  ]
}

Cost Impact

• Cache writes: 1.25x-2x base input price (depending on TTL)
• Cache reads: 0.1x base input price (90% savings)

For typical OpenClaw usage with large system prompts and workspace context that stays constant across turns, this could reduce costs by 60-80% in practice.

Suggested Config

Add config options like:

{
  "agents": {
    "defaults": {
      "models": {
        "openrouter/anthropic/claude-sonnet-4.5": {
          "params": {
            "cacheRetention": "long", // "none" | "short" (5min) | "long" (1h)
            "cacheBreakpoints": ["system", "workspace", "memory"]
          }
        }
      }
    }
  }
}

References

• OpenRouter Prompt Caching: https://openrouter.ai/docs/guides/best-practices/prompt-caching
• Anthropic Prompt Caching: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching

User Impact

This feature would provide immediate cost savings for all users running Anthropic models through OpenRouter, especially those with:

• Large workspace context files
• Long-running sessions with stable system prompts
• Frequent memory/KB searches

/cc @vincentnoca who originally requested this feature