Cost Management

OpenClaw itself is free (MIT licence). What you pay for is model usage: every agent turn, heartbeat, utility task (titles, narration, recaps), embedding and paid tool call goes to a provider that bills per token or draws on a plan quota. This page explains where the tokens go, how to estimate a bill, and which settings reduce it. It is provider-neutral; prices below are examples.

OpenClaw has no built-in spend cap

There is no daily_budget_usd or monthly_budget_usd setting. The config schema is strict, so invented keys like these stop the Gateway from starting (configuration). Set hard limits where billing happens: in your provider's console, on an OpenRouter or ClawRouter key, or by using a subscription plan with fixed quotas.

Where the money goes

Spend	What it is	Main lever
Agent turns	Your messages and the tool loop that follows. Each request resends the system prompt, workspace bootstrap files, tool definitions and session history.	Model choice, /compact, /new, prompt caching
Heartbeats	Scheduled main-session turns, every 30m by default (1h with Anthropic OAuth or Claude CLI auth)	every, activeHours, isolatedSession, lightContext, a cheaper heartbeat.model
Automations	Each cron/automation job is its own agent run	Job schedule and --model
Utility tasks	Session and thread titles, progress narration, Activity recaps	agents.defaults.utilityModel
Other paid features	Media understanding, image and video generation, memory embeddings, web search and fetch, speech	Turn off what you don't use

Upstream lists every feature that can spend a key in API usage and costs, and explains how the prompt is built in Token use.

Runaway patterns

Cause	What happens
Context snowball	A long session grows, and every request carries all of it. Cost per turn rises steadily.
Tool loop	A failing tool appends its error, the model retries, and the history keeps growing.
Large tool output	A huge file listing or web page lands in context and is resent every iteration.
Expensive heartbeat	The main model with full context runs 48 times a day while nothing is happening.
High thinking effort everywhere	Reasoning tokens are billed as output.

Estimate before you pay

Use one formula for any provider:

cost per request = (input_tokens x input_price + output_tokens x output_price) / 1,000,000
monthly cost     = cost per request x requests per day x 30

Cached input is billed at the provider's cache-read price instead of the full input price. Anthropic charges 0.1x the input price for cache reads on most models and 0.05x on Opus 5.5 and Sonnet 5.5.

Model prices used on this page

Per million tokens (input / output), as of 2026-10-18, from Anthropic pricing:

Model	Ref	Input / output	Cache read
Claude Fable 5.1	anthropic/claude-fable-5-1	$10 / $50	$0.25
Claude Opus 5.5	anthropic/claude-opus-5-5	$4 / $20	$0.20
Claude Sonnet 5.5	anthropic/claude-sonnet-5-5	$2 / $10	$0.10
Claude Haiku 5.5 (prompt up to 100K tokens)	anthropic/claude-haiku-5-5	$0.10 / $0.50	$0.01
Claude Haiku 5.5 (prompt over 100K tokens)	anthropic/claude-haiku-5-5	$0.50 / $2.50	$0.05
Legacy: Opus 4.6 to 4.8, Sonnet 4.6, Haiku 4.5		$5 / $25, $3 / $15, $1 / $5	0.1x input

Other providers, per million tokens (input / output), as of 2026-10-18, from each provider's pricing page:

Model	Ref	Input / output	Source
GPT-6 Astra	openai/gpt-6-astra	$10 / $50	OpenAI, Standard tier, short-context rates
GPT-6.1 Sol	openai/gpt-6.1-sol	$2 / $10	same
GPT-6 Luna	openai/gpt-6-luna	$0.10 / $0.50	same
Gemini 3.1 Pro Preview (prompt up to 200K tokens)	google/gemini-3.1-pro-preview	$2 / $12	Google, paid tier
Gemini 4 Argon (not in the 2026.10.4 catalog)		$2 / $10 until 2026-11-30, then $4 / $20	same
DeepSeek V4.1 Flash	deepseek/deepseek-flash	$0.30 / $1.20 at peak, $0.15 / $0.60 off-peak	DeepSeek, cache-miss input
DeepSeek V4 Pro	deepseek/deepseek-v4-pro	$1.32 / $3.96 at peak, $0.66 / $1.98 off-peak	same

OpenAI charges more for long prompts (for example GPT-6.1 Sol $4 / $15). DeepSeek's peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Two dated changes are worth knowing: Gemini 4 Argon, which the 2026.10.10 beta makes OpenClaw's default Gemini model, doubles in price on 2026-12-01; and DeepSeek bills its Search and Fetch APIs at $0.75 per 1,000 requests, which matters if DeepSeek serves your web_search (details in Model Selection). Local models cost nothing per token.

Heartbeat costs

This is the one heartbeat cost method used across clawdocs. Two scenarios:

• Full context: a heartbeat in the main session with history and bootstrap files. Assume 120K input tokens, 500 output tokens, no cache hits.
• Lean: isolatedSession: true and lightContext: true, so no conversation history and no workspace bootstrap files. Assume 10K input tokens, 500 output tokens.

The token counts are assumptions. Measure your own with /usage full on a heartbeat turn, then put your numbers into the formula.

Model	Full context, per beat	Full, 30m (48/day)	Full, 1h (24/day)	Lean, per beat	Lean, 30m	Lean, 1h
Opus 5.5	$0.490	$705.60	$352.80	$0.050	$72.00	$36.00
Sonnet 5.5	$0.245	$352.80	$176.40	$0.025	$36.00	$18.00
Haiku 5.5	$0.061	$88.20	$44.10	$0.00125	$1.80	$0.90
Local (Ollama, llama.cpp)	$0	$0	$0	$0	$0	$0

Worked example, Opus 5.5 full context: (120,000 x $4 + 500 x $20) / 1,000,000 = $0.49 per beat; x 48 x 30 = $705.60 a month. Haiku 5.5 at 120K is above its 100K threshold, so the higher tier applies: (120,000 x $0.50 + 500 x $2.50) / 1,000,000 = $0.061.

Two things move these numbers more than the model does. Isolated, light-context heartbeats cut input by roughly 10x. Restricting activeHours to 14 hours a day removes about 42% of beats. If most of the prompt is a cache hit, the input cost falls further. The flip side: a heartbeat that fires after the cache has expired pays to rewrite it.

Settings that reduce cost

Every block below passes openclaw config validate on 2026.10.4. Merge the parts you want into your config.

1. Route work to the right model

Keep a capable primary model, and send background work somewhere cheaper:

~/.openclaw/openclaw.json

{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-sonnet-5-5",
        fallbacks: ["anthropic/claude-haiku-5-5"],
      },
      // Short internal tasks: titles, progress narration, Activity recaps
      utilityModel: "anthropic/claude-haiku-5-5",
      heartbeat: {
        every: "1h",
        model: "anthropic/claude-haiku-5-5",
        isolatedSession: true,
        lightContext: true,
        activeHours: { start: "08:00", end: "22:00" },
      },
    },
  },
}

• model.primary and model.fallbacks select the chat model; fallbacks are tried in order (Models).
• utilityModel handles short internal tasks. When unset, OpenClaw uses the primary provider's small-model default if it has one, otherwise the primary model.
• heartbeat.model runs heartbeats on a different model from chat. A local model works too, for example "ollama/qwen3:14b" once the Ollama provider is set up (Local Models).
• Per-agent overrides go under agents.entries.<id>.model, never agents.list.

You can make the same changes from the CLI:

openclaw models set anthropic/claude-sonnet-5-5
openclaw models fallbacks add anthropic/claude-haiku-5-5
openclaw config set agents.defaults.heartbeat.every '"1h"'

See the Model Selection Guide for which models suit which tasks.

2. Tune the heartbeat

Change	Effect on heartbeat spend
every: "30m" to "1h"	About -50%
every: "30m" to "2h"	About -75%
activeHours of 14h instead of 24h	About -42%
isolatedSession: true + lightContext: true	Roughly -90% in the scenarios above
Cheaper heartbeat.model	Proportional to the price difference
every: "0m"	Stops the recurring heartbeat. Event-driven wakes, such as a background command finishing, can still run a turn.

Recurring checks (inbox, calendar, health pings) belong in automations, each with its own schedule and model, not in the heartbeat. The old HEARTBEAT.md file is retired; openclaw doctor --fix migrates it into the heartbeat monitor's scratch. See Heartbeat.

3. Keep context small

Habit	Why
/new or /reset when you change topic	Starts a session without the old history
/compact on long sessions	Summarises older messages
Opt into an automatic reset if you want one	The default is no automatic reset. Set session.reset.mode to "daily" (with atHour, default 4) or "idle" (Sessions).
Trim tool output	Ask for summaries, not full dumps
Keep skill descriptions short	The skill list is injected into every prompt

~/.openclaw/openclaw.json

{
  session: {
    reset: { mode: "daily", atHour: 4 },
  },
}

4. Lower thinking effort for routine work

Reasoning tokens are billed as output. Set a lower default and raise it per message when needed with /think high:

~/.openclaw/openclaw.json

{
  agents: {
    defaults: {
      thinkingDefault: "low",
    },
  },
}

Claude Opus 5.5 always thinks: a stored off maps to low (Anthropic provider). On other models off is available. See Thinking levels.

5. Use prompt caching well

OpenClaw sets cache breakpoints for Anthropic and other providers that support caching. cacheRetention accepts "none", "short" (the provider's default window, 5 minutes on Anthropic) or "long" (1 hour on Anthropic). Set it globally, per model, or per agent:

~/.openclaw/openclaw.json

{
  agents: {
    defaults: {
      params: { cacheRetention: "long" },
      // Keep the 1h cache warm across idle gaps
      heartbeat: { every: "55m" },
      // Drop old tool results once the cache has expired
      contextPruning: { mode: "cache-ttl", ttl: "1h" },
    },
  },
}

A 1-hour cache write costs more than a 5-minute one (2x input instead of 1.25x on Anthropic), so "long" pays off only when you come back within the hour. See Prompt caching.

6. Use a subscription you already have

Plan	How OpenClaw uses it	Caveat
Claude Pro or Max	Claude CLI auth (claude-cli/* route), reusing your Claude Code login	Usage draws on your plan's limits, and Anthropic's terms for this can change. API-key billing is more predictable for shared automation (Anthropic provider).
ChatGPT plans	Sign in with Codex OAuth (openai/gpt-6-astra by default)	An OpenAI API key is separate pay-as-you-go billing (OpenAI models)
Google One AI plans	Not usable as Gemini API credit	Gemini API keys are billed separately

Track what you spend

Where	What it shows
/status in chat	Session tokens and estimated cost (API-key models)
/usage full in chat	Usage footer on every reply
/usage cost in chat	Cost summary from OpenClaw's session logs
openclaw gateway usage-cost --days 7	The same summary from the CLI. Add --all-agents or --agent <id>, and --json for scripts.
openclaw status --usage	Provider-reported quota and usage snapshots
Control UI, Usage	Session-derived estimates, plus provider-reported spend when an admin key is set

OpenClaw's own figures are estimates from session logs. For actual billing, set ANTHROPIC_ADMIN_KEY or OPENAI_ADMIN_KEY; the Control UI then shows 30 days of provider-reported cost alongside them (Usage tracking).

Hard limits, at the provider

Where	How
Anthropic Console	Workspace spend limits and alerts
OpenAI dashboard	Project budgets and usage alerts
OpenRouter	Per-key credit limits. Model ref openrouter/auto lets OpenRouter pick the model per prompt (OpenRouter provider).
ClawRouter	Monthly budget per key, shown in openclaw status --usage (ClawRouter)

Community dashboards

These are third-party projects, not reviewed by clawdocs. Check what they read and where they send it before connecting them to your Gateway.

Tool	What it does
openclaw-dashboard	Usage-limit notifications and cost analysis by model
ClawWatcher	Token usage and cost per model
OpenClaw Cost Calculator	Pre-deployment estimates

Quick start: cut costs today

1. Run /usage full for a day and openclaw gateway usage-cost --days 1 to see where tokens go.
2. Make heartbeats lean: isolatedSession: true, lightContext: true, activeHours, and every: "1h" or longer.
3. Put heartbeats and utilityModel on a cheap model (Haiku 5.5 or a local model).
4. Start new sessions with /new when the topic changes, and /compact long ones.
5. Lower thinkingDefault and raise it per message when needed.
6. Set a spend limit in your provider's console.

See also

• Model Selection: which model for which task
• Local Models: zero per-token cost
• Cloud GPU & Self-Hosted Models: renting GPUs and inference APIs
• Performance Tuning: latency and resource tuning
• Monitoring: logs, health and metrics