OpenClaw itself is free. The API costs to run it are not.
Without configuration, OpenClaw routes all requests to your default model — which is often a premium model that's overkill for most tasks. It's like using a chef's knife to open a package: technically works, costs more than necessary.
Here's how to actually reduce your monthly API spend.
Why Costs Get Out of Hand
Default model selection. If you configure OpenClaw with a GPT-4 or Claude Opus key and don't set up routing, every request — including "what time is it?" and simple calendar queries — goes through the premium model.
Skill verbosity. Some skills send enormous context windows with every request. A poorly written email skill might send your last 100 emails as context for every email-related query, even when only the last 5 are relevant.
No request batching. Without configuration, your agent processes each event as it arrives. High-volume inputs (active Slack channels, busy email inboxes) create proportionally high API costs.
Debug mode left on. Some developers enable verbose logging during setup. Verbose modes increase context window usage significantly. Turn them off in production.
Model Routing: The Most Impactful Change
Model routing lets you assign different tasks to different models — expensive models for complex tasks, cheap models for simple ones.
OpenClaw's model configuration lives in ~/.openclaw/openclaw.json (JSON5 format). The model failover system covers both routing and fallback behavior.
The basic routing pattern:
Define your primary model and fallbacks in your config:
{
agents: {
defaults: {
model: {
primary: "anthropic/claude-sonnet-4-6", // main model
fallback: "anthropic/claude-haiku-4-5", // cheaper fallback
}
}
}
}
For per-agent model assignment, configure each agent in the agents.list array with its own model settings. The concept applies universally: identify your task categories, assign appropriate models, and don't use expensive models for tasks that don't need them.
Typical cost reduction from routing: 40–70%, depending on your usage patterns. If most of your agent's work is simple queries and status checks, the savings are larger. If most of your work is complex synthesis and drafting, the savings are smaller.
Auth Profile Rotation for Multi-Key Setups
If you have multiple API keys — for example, both Anthropic and OpenAI accounts — OpenClaw supports auth profile rotation. This lets you:
- Use one provider as primary, another as failover
- Rotate across keys to stay within rate limits
- Switch providers based on cost or performance
Configure auth profiles in your openclaw.json:
{
agents: {
defaults: {
model: {
primary: "anthropic/claude-sonnet-4-6",
fallback: "openai/gpt-5.2", // fallback to different provider
}
}
}
}
Auth profiles for each provider are stored in ~/.openclaw/agents/main/agent/auth-profiles.json.
Why this matters for cost: Different providers have different pricing structures for different task types. During high-traffic periods, routing overflow to a less expensive provider can meaningfully reduce costs.
Skill-Level Cost Controls
Each skill that makes API calls adds to your costs. Control this at the skill level:
Limit context window per skill. Most skills accept a max_context_tokens configuration. Set this to the minimum needed for the skill to work, not the maximum allowed by the model.
Reduce polling frequency. A skill checking your inbox every 30 seconds costs twice as much in webhook/polling API calls as one checking every 60 seconds. For most use cases, the latency difference is unnoticeable.
Disable skills you don't actively use. A skill that's installed but idle can still consume API calls for background checks. If you haven't used a skill in two weeks, uninstall it.
Right-Sizing for Your Usage Pattern
High-volume communication users (active Slack, busy email inbox): model routing is critical. Without it, volume overwhelms costs quickly. Route simple message triage to the cheapest available model.
Light users (mainly scheduling and basic queries): costs are unlikely to be a problem. Focus on getting the setup working correctly before optimizing.
Automation-heavy users (scheduled tasks, monitoring): evaluate which tasks need model intelligence and which can be handled by simple logic. Some scheduled tasks don't need a language model at all — they're just queries to external APIs.
Monitoring Your Actual Usage
Before optimizing, know where your costs are coming from:
Provider dashboards. Anthropic and OpenAI both provide usage breakdowns by model and time period. Check weekly when you're first setting things up.
OpenClaw usage logs. The agent logs include request counts per skill and model. openclaw logs --usage (check your version's flag) shows aggregated usage.
Set budget alerts. Both Anthropic and OpenAI support spend alerts. Set one at 120% of your expected monthly spend so you get a warning before it becomes a surprise.
What to Expect
With model routing configured:
- Simple tasks (status queries, basic calendar checks): €0.001–0.003 per request
- Moderate tasks (email summarization, skill execution): €0.01–0.05 per request
- Complex tasks (drafting, analysis, synthesis): €0.05–0.20 per request
A typical solo user with moderate usage: €5–25/month in API costs after optimization. Heavy users with active automation: €30–80/month.
These are self-hosted API costs — what you'd pay running OpenClaw yourself. With Volos, API costs are bundled into usage credits ($15/month included, top-ups available) so you don't manage API keys at all.
