What it costs to run OpenClaw

As of September 30, 2026, OpenClaw has a $0 software license and no first-party paid tier or hosted service.

Models, external tools, hosting, storage, and human time can still cost money. We built a cost worksheet that keeps those bills and missing inputs separate.

The OpenClaw FAQ and MIT license establish the software fee. Our worked example separates sourced model rates from assumed quantities and other costs.

Free software still has operating costs

OpenClaw runs a Gateway on your own infrastructure, which can be a local computer or a VPS. Its FAQ describes that arrangement. A company offering a managed OpenClaw installation can charge for hosting or support separately; the software license does not set that company's price.

Start with the configuration you intend to use. List the accounts behind the model, search tool, media service, and host. Then check what each account includes and what happens after the allowance runs out.

Table 1 Download CSV

Component	What belongs in the estimate	What to check
OpenClaw software	$0 license fee	MIT license; separately priced hosting or support
Model subscription or prepaid credits	The amount you allocate to this period and setup	Exact account, permitted authentication route, allowance, and excess-usage rules
Model API use	Chargeable input, output, cache reads, and cache writes	Exact model, provider, tier, region, and price per million tokens
Search, browser, extraction, and other tools	Configured services' uncovered charges	Included calls or credits, unit prices, and the account that pays
Media and embeddings	Paid image, video, audio, or embedding calls if used	Provider and billing unit; local processing where configured
Computer and storage	Host charge or machine allocation, storage, and electricity	Your actual plan, equipment, retention, and operating period
Operator time	Setup, maintenance, and intervention minutes	Your chosen hourly value, kept separate from cash payments
Tax and fees	Applicable charges you can establish	Invoice or checkout details; leave missing amounts unknown

OpenClaw's API usage cost guide documents model calls and configured external services that can generate charges. A tool call may consume model tokens and incur a separate service charge. For search, the search documentation specifies the selected provider's normal account limits and billing.

For model selection, use current model API rates and the agent model rankings to choose candidates. Those pages do not establish what a particular OpenClaw configuration will spend or how reliably it will finish your tasks.

Your authentication route determines who pays

Bring your own API key means the configured API account can receive usage charges. Holding a consumer subscription with the same provider does not establish that those API calls are included.

OpenClaw also documents an official Claude CLI route, using the Claude CLI runtime for the selected model. Its provider plugin guide describes this setup. On the upstream side, Anthropic's June 15 update says Claude CLI and Agent SDK usage still draw from subscription limits. The monthly-credit changes described farther down that article were paused. Its historical table is not a current credit offer.

There is a concrete billing trap here. Anthropic says an ANTHROPIC_API_KEY environment variable takes precedence over subscription authentication in Claude Code and produces API charges. Check that variable and the actual OpenClaw runtime before budgeting a session as included usage. Claude Code plan documentation also explains shared plan limits and the option to continue with paid API credits.

These are specific documented routes. Anthropic's authentication policy distinguishes API access from authorized subscription access. A successful login alone cannot establish that every third-party tool can consume a plan's allowance.

Local inference still uses a computer

OpenClaw's custom provider documentation supports local backends including Ollama, LM Studio, vLLM, SGLang, and llama.cpp. That can move model inference onto equipment you supply.

This provider documentation defaults omitted token-cost metadata to zero. That zero is a configuration value. It does not measure the machine, electricity, storage, or maintenance cost.

An existing computer may have no new purchase attached to a trial. You can record that marginal hardware expense as an explicit zero while separately allocating equipment cost or operator time for a fuller budget. Label which view you are using. The self-host calculator helps explore hardware assumptions. It cannot quote a complete OpenClaw operating bill.

For the broader hosting and workflow decision, the OpenClaw versus Hermes Agent guide compares the documented setups. OpenClaw alternatives covers other agent options.

Credits cannot be counted twice

Download the OpenClaw cost worksheet. The XLSX has a blank estimate, a billing-period ledger, instructions, and a separate worked example. The reader inputs begin blank. Model quantities use millions of tokens: enter six million input tokens as 6.

Choose one billing period and one reporting currency. Then price each relevant row or explicitly mark it unused. A usage row confirmed as covered by a plan or prepaid balance adds no new cash charge. Put the applicable plan or credit allocation in its own row. Uncovered usage needs its own estimate.

Calculate the estimate as follows:

model estimate = input millions × input rate
               + output millions × output rate
               + separately priced cache reads and writes

operator value = minutes ÷ 60 × hourly value

operating estimate = allocated expenses + expected additional charges
                   + operator value, if supplied

full total = unknown while a relevant component is unresolved

Use the model cost calculator to estimate inference. Hosting, external tools, storage, electricity, tax, and operator time still need their own inputs.

Keep these four ledger amounts separate:

Table 2 Download CSV

Ledger field	What it records	Example of the distinction
Amount already paid	Cash transferred to a provider	Buying a prepaid balance
Allocated recurring expense	The portion assigned to this setup and period	A share of an existing subscription or host plan
Estimated additional amount	An expected uncovered charge	API use after included credit is exhausted
Verified invoice amount	A provider charge confirmed for the stated scope	The finished period's billed usage

Buying credit and later consuming that same credit are not two cash expenses. A subscription's included allowance also cannot be converted into an extra cash charge simply by multiplying its usage by an API rate. Use the provider's actual rules for that account.

Our worksheet records the component, account and authentication route, period, currency, quantity and unit, rate and basis, credit treatment, source URL, check date, and unknown reason. It excludes mismatched currencies from the selected currency's known subtotal and leaves the full total blank. Supply an explicit allocation before turning an annual plan into a monthly amount.

Assumed model use costs $18

Here is a concrete example of the arithmetic. None of the workload or allocation figures below came from an OpenClaw run or invoice.

Assume 200 total model responses over 30 days, averaging 30,000 input tokens and 3,000 output tokens each. That gives six million input tokens and 0.6 million output tokens. The 200 includes any model responses used for retries, subagents, or compaction. That does not mean 200 successful jobs.

For one sourced rate, Claude Sonnet 5.5's exact API identifier is claude-sonnet-5-5. Its model page lists $2 per million input tokens and $10 per million output tokens. These USD rates were checked September 30, 2026. The example uses the global Standard API route with uncached input. It assumes no Batch discount, cache charges, regional premium, or provider-side extra tool charge. Check the pricing conditions for your route.

Table 3 Download CSV

Example component	Quantity or allocation assumption	Rate or basis	Estimated USD
OpenClaw software	One period	Sourced $0 software fee	$0
Model input	6 million tokens	Sourced $2 per million	$12
Model output	0.6 million tokens	Sourced $10 per million	$6
External tools	One 30-day allocation	Assumed $8 budget placeholder	$8
Host compute	One 30-day allocation	Assumed $10 budget placeholder	$10
Storage	One 30-day allocation	Assumed $2 budget placeholder	$2
Subscription allocation	Not used in this API-key example	Explicit $0 assumption	$0
Operator time	Assumed 90 minutes	Assumed $60 per hour	$90
Tax and electricity/hardware	Not supplied	Unknown	Unknown

Model use costs 6 × $2 + 0.6 × $10 = $18. Adding the assumed tools, host, and storage gives $38 of listed cash components. Valuing the assumed operator time adds 90 ÷ 60 × $60 = $90, giving $128 of listed operating components.

A full total remains unknown because tax and electricity/hardware were not supplied. Host, tool, storage, and time figures are placeholders to replace with your own budget inputs. They are not vendor quotes or typical OpenClaw costs.

Input tokens include more than what you type. OpenClaw's token-use guide lists system instructions, conversation history, tool results, attachments, and compaction context. Its subagent documentation says each subagent has its own context and token usage. Count billed model activity across the setup before applying a rate.

Your final bill needs provider receipts

OpenClaw's local usage history can help explain activity, but a local estimate and a provider charge have different scopes. Its usage-tracking guide separates local session accounting from provider usage and billing records. Some provider billing cards can show account-level history when the required credentials are configured.

For a finished period:

1. Match the provider account, model route, start and end dates, and billing timezone.
2. Save provider usage and charge records, plus receipts for hosting, tools, and storage. Note included allowances, prepaid balances, and discounts.
3. Reconcile input, output, and cache quantities across sessions and subagents. Keep missing or canceled-call usage unknown until the provider supplies it.
4. Replace an estimate only with a verified charge covering the same scope. Keep tax, operator time, and unrelated account usage separate.

OpenClaw's API cost guide says local cost displays depend on available usage information and pricing metadata. A token count or remaining plan quota by itself is not a dollar invoice.

Before committing to the setup, identify the account that pays for the model and the services it calls. Fill those rows first, then price the computer and your time. Leave the total unfinished until the missing amounts are resolved.

Questions

Can OpenClaw use an existing Claude subscription?

Yes, through OpenClaw's documented Claude CLI route when your Claude account is eligible. Anthropic's current update says Claude CLI usage still draws from subscription limits. An ANTHROPIC_API_KEY environment variable can instead trigger API charges. Check the exact runtime and account before treating usage as included in a plan.

Are tool charges included in a model bill?

Not always. A model bill can include tokens produced while calling tools, while search, extraction, media, or other services may charge through separate accounts. OpenClaw's configured providers determine the route. Record each tool's allowance and uncovered charge, and leave its cost unknown when billing information is missing.

How should I budget for a local model?

Choose a billing period, then record the machine allocation, electricity, storage, and operator time alongside any paid external services. OpenClaw supports local model backends, but zero token-price metadata does not establish a zero operating bill. The self-host calculator can help explore hardware assumptions. It does not supply an OpenClaw invoice.