OpenClaw itself is free. The bill comes from the models you route through it. How much that adds up to depends almost entirely on two things: which model you pick, and how often you call it.

This post walks through what a real monthly bill looks like for different use cases, plus a full pricing comparison so you can estimate your own.

What OpenClaw actually charges

Zero. Install is free, the CLI is free, the config is free. Every dollar you spend is going directly to the LLM provider you’ve routed through: OpenAI, Anthropic, Google, haimaker, whoever. OpenClaw is just the orchestration layer.

The good news: because OpenClaw works with any OpenAI-compatible API, you have total control over your costs. Swap models in a config file and watch the bill change. Use local models for $0/token. Route batch work to a cheap provider and interactive work to a premium one.

Typical monthly bills by use case

These are rough ranges from actual users, assuming a developer working 20 days a month.

Light usage (~2 hours/day, mostly chat and small edits)

ModelRough Monthly Bill
MiniMax M2.5$2–$8
Qwen3 Coder$4–$12
Gemini 3 Flash$5–$15
GPT-5.4 Mini$10–$25
Claude Sonnet 4.6$25–$60
Local (Ollama)$0 (after hardware)

Daily driver (~6 hours/day, active coding, tool calling, multi-file edits)

ModelRough Monthly Bill
MiniMax M2.5$5–$20
Qwen3 Coder$15–$40
GPT-5.4 Mini$30–$80
Gemini 3.1 Pro$40–$100
Claude Sonnet 4.6$80–$200
GPT-5.4$100–$250

Heavy agent usage (long autonomous runs, large contexts, many parallel sessions)

ModelRough Monthly Bill
MiniMax M2.5$30–$100
Qwen3 Coder Plus$80–$250
Claude Sonnet 4.6$300–$800
GPT-5.4$400–$1000
Claude Opus 4.6 / GPT-5.4 Pro$1500–$5000+

These are not scientific numbers. Your mileage will vary based on context window size, how often you’re iterating, how verbose the model’s output is, and whether you’re caching anything.

Full pricing table

Current per-million-token pricing across OpenClaw-compatible models, sorted roughly by value:

ModelInput ($/M)Output ($/M)ContextNotes
Qwen3.5-Flash$0.07$0.261MCheapest paid path
MiniMax M2.5$0.12$1.00196KBest cheap default
Qwen3 Coder$0.22$1.00262KCheap + coding-tuned
GPT-5.4 Nano$0.20$1.25400KBudget OpenAI
Grok 4.1 Fast$0.20$0.502MCheapest huge context
Grok Code Fast$0.20$1.50256KCoding specialist
MiniMax M2.5 Lightning$0.30$2.401MFast + long context
DeepSeek V3.2$0.28$0.40164KCheapest output tokens
Gemini 3 Flash$0.50$3.001MFree tier available
GPT-5.4 Mini$0.75$4.50400KBest OpenAI value
Qwen3 Coder Plus$0.65$3.251MMid-tier coding
Gemini 3.1 Pro$2.00$12.001MStrong reasoning
Grok 4.20$2.00$6.002MNew reasoning model
GPT-5.4$2.50$15.001.05MOpenAI flagship
Claude Sonnet 4.6$3.00$15.00200KPremium coding
Grok 4$3.00$15.00256KReal-time web search
GPT-5.4 Pro$30.00$180.001.05MResearch-grade only
Claude Opus 4.6$15.00$75.00200KHighest-quality coding

Gemini’s free tier (1,000 requests/day on 2.0 Flash) is worth flagging. If your usage fits inside that limit, it’s literally zero dollars per month.

Three real ways to cut your bill

If you’re spending more than you’d like, there are three levers that actually work.

1. Route most traffic to a budget model

This is by far the biggest lever. The cost gap between MiniMax M2.5 and Claude Sonnet 4.6 is 25x on input and 15x on output. For most coding work (refactoring, file edits, documentation, explaining code, tool calls), the quality gap between them is much smaller than the price gap.

The pattern that works: set a cheap model as your default and only escalate to a premium model when the cheap one gets stuck. In practice, 70–80% of OpenClaw requests don’t need a flagship model. You notice almost no quality drop on daily work and your bill drops by 60–90%.

See the cheapest models guide for the specific routing pattern.

2. Trim your context window

Every file you load into context costs money on the input side. Most people over-load context because it’s easy: loading the whole repo is less effort than picking the right three files. But on a big codebase, the difference between 20K and 200K of input tokens is an order of magnitude on the bill.

OpenClaw’s /files and selective file loading exist for exactly this reason. Use them.

3. Pick models with cheaper output tokens

Output tokens cost more than input tokens across every provider, usually 3–6x more. On code generation tasks, output dominates the bill. A model like DeepSeek V3.2 ($0.40/M output) vs. Claude Sonnet 4.6 ($15/M output) is a 37x ratio on the expensive side of the equation.

For high-output workflows (large refactors, code generation, document drafting), pick a model with a cheap output tier even if input is slightly more expensive. For context-heavy workflows with small outputs (code review, summarization), pick a model with cheap input.

How to estimate your own bill

If you want an actual number before you commit, here’s the rough math:

  1. Estimate input tokens per session: roughly 1K tokens per file loaded + your prompt + the conversation history. A typical OpenClaw session with 10 files loaded and a few back-and-forth turns lands around 20–50K input tokens.
  2. Estimate output tokens per session: 2–10K is typical for coding work. Agent-style runs with long autonomous loops can push 50–100K+.
  3. Multiply by sessions per day × 20 working days.
  4. Plug into the pricing table above.

Example: 5 sessions/day, 30K input + 5K output each, 20 days, on GPT-5.4 Mini.

  • Input: 5 × 30K × 20 = 3M tokens × $0.75 = $2.25
  • Output: 5 × 5K × 20 = 500K tokens × $4.50 = $2.25
  • Total: ~$4.50/month

Same usage on Claude Sonnet 4.6:

  • Input: 3M × $3 = $9
  • Output: 500K × $15 = $7.50
  • Total: ~$16.50/month

Same usage on MiniMax M2.5:

  • Input: 3M × $0.12 = $0.36
  • Output: 500K × $1 = $0.50
  • Total: ~$0.86/month

Those numbers are small because the example is a light usage pattern. Multiply by 5–10x for a daily driver, by 20–50x for heavy agent workflows.

The simplest cost-control setup

If you just want one setup that gets you most of the savings without thinking too hard:

  1. Route everything through haimaker.ai so you only manage one API key.
  2. Set MiniMax M2.5 as your default model in OpenClaw.
  3. Keep Claude Sonnet 4.6 or GPT-5.4 as a backup for when M2.5 gets stuck on something hard.
  4. Swap models with /model mid-session when you need the upgrade.

That’s it. Most people end up spending 10–30% of what they were paying on a single-model setup, with no noticeable loss in quality on day-to-day work.