OpenClaw Cost Optimizer

Analyzes the user's OpenClaw config and rewrites it to minimize Anthropic API spend. Works on YAML and JSON configs. Reads pricing tables from references/pricing.md before starting.

Step 0 — Load pricing reference

Read references/pricing.md now. It contains current model rates and the savings formulas used for estimates. Do this before any analysis.

Step 1 — Locate the config

Check these locations in order, stop at the first match:

1. Path the user explicitly provides
2. ./openclaw.yaml or ./openclaw.json in the current working directory
3. ~/.openclaw/config.yaml
4. ~/.openclaw/openclaw.yaml
5. ~/openclaw.yaml

If none found, ask the user to paste their config directly.

Step 2 — Audit: the five cost levers

Work through every lever. Flag each issue with severity HIGH / MEDIUM / LOW.

Lever 1 — Prompt caching (HIGH impact, ~60–90% input cost reduction)

Flag HIGH if any of these are missing for Anthropic models:

• cacheRetention not set, or set to "none"

Fix: Add cacheRetention: "long" for the primary model. Use "short" (5 min) for agents that handle fast-changing context, "long" (1 hr) for everything else.

agents:
  defaults:
    models:
      anthropic/claude-opus-4-6:
        params:
          cacheRetention: "long"

Note: caching is API-key only. Subscription setup-tokens do not honor it.

Lever 2 — Model routing (HIGH impact)

Flag HIGH if: a single model (especially Opus) is used for ALL agents with no per-agent overrides.

Fix: Route by task complexity:

agents:
  list:
    - id: planner
      params:
        model: anthropic/claude-opus-4-6
    - id: coder
      params:
        model: anthropic/claude-sonnet-4-6
    - id: reviewer
      params:
        model: anthropic/claude-sonnet-4-6
    - id: triage
      params:
        model: anthropic/claude-haiku-4-5

Lever 3 — Thinking level (MEDIUM impact)

Flag MEDIUM if: thinking: "adaptive" or thinking: "high" is set as the default for ALL agents, including simple ones.

Fix: Scope thinking: "adaptive" only to the primary reasoning agent. Set thinking: "low" or omit for utility agents.

agents:
  defaults:
    models:
      anthropic/claude-opus-4-6:
        params:
          thinking: "low"   # default: low
  list:
    - id: planner
      params:
        thinking: "adaptive"  # override only where needed

Lever 4 — 1M context window (MEDIUM impact)

Flag MEDIUM if: context1m: true is set globally or for agents that don't need it.

The 1M context beta header triggers surcharge pricing for Anthropic API calls. Most agents don't need it.

Fix: Remove context1m: true from the global defaults. Re-add only to agents that demonstrably require >200K tokens of context.

Lever 5 — Fast mode (LOW impact, but easy win)

Flag LOW if: fastMode is not enabled for Sonnet agents doing quick tasks.

Fix: Add fastMode: true to Sonnet model params for agents doing rapid back-and-forth tasks (search, triage, review). This reduces latency and can lower costs on high-frequency agents.

models:
  anthropic/claude-sonnet-4-6:
    params:
      fastMode: true
      cacheRetention: "long"

Step 3 — Generate the optimized config

Rewrite the full config applying all fixes. Preserve every key the user had that is not being changed. Add comments above each changed block explaining what changed and why.

Format:

# COST OPTIMIZER — changed blocks annotated below
# Original: \x3Cwhat it was>  →  Optimized: \x3Cwhat it is now>

Step 4 — Show the diff and cost estimate

Before writing anything, show the user:

1. Issues found — list each one with severity and one-line description
2. Estimated monthly savings — use the formula from references/pricing.md with assumed 500K tokens/day baseline
3. Diff — show only the changed blocks (not the full file unless it's short)

Example output format:

Issues found (3):
  [HIGH]   No prompt caching — estimated +$340/mo at current usage
  [HIGH]   All agents using Opus — Haiku suitable for triage/review agents
  [MEDIUM] thinking: adaptive set globally — limits to reasoning agent only

Estimated savings: ~$290/mo (68% reduction)
  Before: ~$430/mo  →  After: ~$140/mo

Changed blocks:
  + agents.defaults.models.anthropic/claude-opus-4-6.params.cacheRetention: "long"
  + agents.list[triage].params.model: anthropic/claude-haiku-4-5
  ~ agents.list[planner].params.thinking: adaptive → low (default); kept adaptive on planner only

Step 5 — Confirm and write

Ask the user: "Apply these changes to your config?"

• If yes: write the optimized config back to the same file path. Then tell the user to run openclaw gateway restart (or it will auto-restart if config watch is enabled).
• If no / partial: apply only confirmed changes.
• If they want to review first: show the full optimized YAML before writing.

Edge cases

• Subscription setup-token auth: Note that cacheRetention has no effect with subscription tokens — it's API-key only. Tell the user if their auth config uses sk-ant-oat-* tokens.
• Multiple agents with no IDs: If the config has unnamed agents, suggest adding IDs to enable per-agent model routing.
• Non-Anthropic providers: Only audit Anthropic model blocks. Leave OpenAI, Gemini, and other provider blocks untouched.
• Config not found: If no config file is found and no paste provided, offer to generate a fresh cost-optimized starter config instead.