How much does OpenClaw cost per month? A budget in tokens

An always-on OpenClaw assistant costs whatever its tokens cost: roughly steps per day × tokens per step × 30 days, priced by the model you pick. With the workload below, a month is tens of millions of input tokens and only a few million output tokens, so the model price decides most of the bill. Here is how to get your own number in five minutes.

You set the assistant up, it answers in your chat channels all day, and the first question is always the same: what will this be at the end of the month? Nobody can give you a universal figure, because OpenClaw bills you for whatever it sends to the model, and that depends on your setup. So let's measure the pieces.

What OpenClaw sends on every step#

The OpenClaw docs (token-use reference, checked 26 October 2026 against version 2026.9.3; the page carries no publication date) say the system prompt is rebuilt on every run. It contains the tool list, skills metadata, workspace bootstrap files such as AGENTS.md and MEMORY.md, the current time and runtime metadata. Injected files are capped by agents.defaults.bootstrapMaxChars and bootstrapTotalMaxChars.

On top of that, everything the model receives counts toward context: history, tool calls and their results, attachments and compaction summaries. So a step is never "one short question". It is the whole prompt plus the whole conversation so far, and that is why input tokens dwarf output tokens. If the idea of resending history is new to you, why a long chat costs more explains it with a small example.

Two details from the same page matter for budgeting. DMs share the agent's main session, while each group chat gets its own session, so a busy group grows its own history. And heartbeat turns can keep the provider cache warm across idle gaps, but each heartbeat turn still costs tokens: an idle assistant is not a free assistant.

A daily workload you can change#

There is no measured "typical OpenClaw day" to quote, so take an assumption and replace the numbers with yours. Say your assistant makes 150 model calls a day (messages, tool steps, a few scheduled checks). Say each call sends about 12,000 input tokens (system prompt, history, tool output) and gets back about 600 output tokens.

• Per day: 150 × 12,000 = 1,800,000 input tokens and 150 × 600 = 90,000 output tokens.
• Per month (30 days): 1,800,000 × 30 = 54,000,000 input and 90,000 × 30 = 2,700,000 output.

Notice the shape: input is twenty times output. That is the normal shape for an agent, and it is the reason a model's input price matters more here than in a chat where you type short questions and read long answers. For the difference between the two sides, see input vs output tokens.

What that month costs on three models#

Here is the same month, 54,000,000 input and 2,700,000 output tokens, priced live for the three models you can use through CheapTokens. Input and output are counted together, because one request pays for both. In total, each of them comes out cheaper than OpenRouter, and the table shows both columns side by side.

In numbers: DeepSeek-V4-Flash for the month is $1.08, GLM-5.3-Flash is $1.19, and MiniMax-M2.7 is $3.11. The prices behind them are on the pricing page and update with the table.

Which one to pick? DeepSeek-V4-Flash gives you a 380K context window and up to 32,768 output tokens per answer at one price around the clock, which suits long agent sessions. GLM-5.3-Flash has a 390K context and up to 8,192 output tokens, a good fit for routine and exploratory steps. MiniMax-M2.7 has 180K context and up to 8,192 output tokens. A full comparison by task lives in the cheapest model for OpenClaw, so I won't repeat it here.

Watch the bill from inside OpenClaw#

OpenClaw can show you what you spend while you spend it. Per the docs, /status shows the session model, context usage, the input and output tokens of the last response and an estimated cost. /usage off|tokens|full adds a footer to every response, and /usage cost summarizes cost from the session logs.

A small habit that works: run /usage tokens for a day. After a few dozen replies you will see your real input per step, and you can put it into the formula above instead of my 12,000.

One caveat from the docs. The cost estimate comes from the cost values you set per model in models.providers.<provider>.models[] (USD per 1M tokens: input, output, cache read, cache write). If you leave them out or copy old numbers, the estimate is off, while amounts billed by the provider take precedence over the catalog estimate. Set them from the current table, then compare /status with your balance after a day.

Four habits that shrink the month#

The docs list the ways to reduce pressure on context. Here they are with what they do to the arithmetic.

1. Run /compact on long sessions. It summarizes the history, so the next steps resend a summary instead of everything. Say it takes the average input per step from 12,000 to 8,000.
2. Trim large tool outputs. A tool that returns a whole file or a page of logs is resent on every following step. Cut it before it enters the session.
3. Shrink images. For screenshot-heavy sessions, lower imageMaxDimensionPx.
4. Keep skill descriptions short and use a smaller model for verbose exploration.

Here is what habit 1 does to the same assistant. With 8,000 input tokens per step, a month is 150 × 8,000 × 30 = 36,000,000 input and still 2,700,000 output. On DeepSeek-V4-Flash that is $0.81 instead of $1.08. The answers are the same length, so the saving comes entirely from the input side.

Set a budget you can't overshoot#

Work backwards. Decide what you are ready to spend, top up the balance by that amount, and treat the prepaid balance as the ceiling: when it runs out, the requests stop. Then check your real usage after the first day with /status, multiply by 30, and compare with the table above.

A simple routine:

1. Pick the model from the table that fits your task.
2. Run a normal day with /usage tokens on.
3. Replace 150 and 12,000 in the formula with what you saw.
4. Re-read the ::cost table with your numbers and top up accordingly.

If you haven't connected OpenClaw yet, how to connect OpenClaw to CheapTokens walks through the config in ~/.openclaw/openclaw.json, and the docs have the API address and model ids. For a wider look at agent budgets, see how to estimate the cost of an AI agent.