What OpenClaw really costs per month: server, models and the hidden lines

Ask a search engine how much OpenClaw costs per month and the answers span three orders of magnitude: Hostinger’s guide says “approximately $6 to $200+ per month”, Kilo’s says $0 to $520 and up. Ask an AI assistant for the cheapest way to run it around the clock and you get a server price with a footnote that “API usage is billed separately”. Both are right, and neither is the number you will pay. The software itself is free: OpenClaw is MIT-licensed, and its site says “No subscription. No hosted tier. No token.” The bill is everything around it, and it has three lines:

server + models + your time.

This page prices each line from the sources themselves, read on 20 October 2026: the server from the providers’ own price pages, the models from OpenRouter’s price list and OpenClaw’s own documentation of what a turn sends, the hidden lines from the same docs. The model line is worked out in tokens first, so the arithmetic survives the next price change. Aplite is one of the scenarios below, and this page is written by the company behind it.

Seven ways to run it, one month each

Server prices are per month as printed, before tax unless the page says otherwise. The model column assumes the example worked out in the model line: thirty messages a day on a cheap or a frontier model. “Your time” is what you do yourself.

The rest of the page is each line in detail, with the quotes.

The server line

Your own machine. The upstream FAQ puts the floor at “1 vCPU, 1 GB RAM, ~500 MB disk” and recommends 2 GB or more; a laptop’s advantage is listed as “no server cost”, its disadvantages as “sleep/network drops disconnect it, OS updates/reboots interrupt it, must stay awake”. The electricity is real but small on a small machine: Apple rates the Mac mini M4 at 4 W idle and 65 W maximum, which over a 720-hour month is 2.9 to 47 kWh. At the US residential average of 18.31 ¢/kWh (July 2026) that is $0.53 to $8.57; at Germany’s €0.3869 (second half of 2025) it is €1.11 to €18.11, and at the EU average of €0.2896, €0.83 to €13.55. A desktop with a GPU is a different bill: NVIDIA rates the RTX 5090 at 575 W of graphics power alone, which would be 414 kWh, about $76 or €160, if it ran flat out all month; it will not, but the idle draw of the whole box is what you pay for 24/7.

Running the model locally is what makes this scenario $0, and it is where the upstream FAQ is blunt: asked whether local models are a good fit, it answers “Usually no. OpenClaw needs large context + strong safety; small cards truncate context and skip provider-side safety filters”, adds that “smaller/quantized models raise prompt-injection risk”, and says to run “the largest model build you can locally”. Size is memory: Ollama’s Gemma 4 weighs 6.6 to 9.5 GB at the E4B size, 16 to 19 GB at 26B and 19 to 20 GB at 31B, before the context window.

Oracle Always Free. The allowance is “2 OCPUs and 12 GB of memory” on the Arm shape, or two AMD micro instances with 1 GB each, which is exactly the floor. The clause that ends it: “Idle Always Free compute instances may be reclaimed by Oracle” when, over seven days, CPU, network and memory all stay under 20 %. An agent that waits for messages is idle by that definition. The other free clouds and their expiry dates are in OpenClaw hosting for free.

A VPS. The three the assistants name most, as printed on 20 October 2026: Hostinger KVM 1 (1 vCPU, 4 GB) at $6.49 a month, “paid upfront”, “Renews at $11.99/mo for 2 years”; DigitalOcean’s Basic Droplet at $6.00 for 1 GiB and $12.00 for 2 GiB; Contabo Cloud VPS 4 (4 vCPU, 8 GB) at €5.50 including VAT “for the first 24 months”. Hetzner, the fourth name, renders its prices only in the browser; its cost-optimized page listed the entry type CX23 at 2 vCPU and 4 GB and marked every type “currently unavailable” on the day we read it, and its docs add €0.50 a month for a public IPv4 address, which the agent needs to be reached. Backups, snapshots and a domain are the small lines on top.

A managed host. Someone else runs the server and the install. The seventeen services that sell this, with price from, free plan or trial, whether the key is yours and where the servers are, fill Managed OpenClaw hosting compared; the cost question splits them in two. Hosts where the model account is yours, so the hosting price is the whole of their bill: Flux from $5.49, Railway from $5, OneClickClaw €14.99, OpenClaw Cloud $19, BetterClaw $19, xCloud $19.99 after the first month, MyClaw $29, ClawTrust $29. Hosts that include some model usage: Runware at $0 for hosting with tokens billed per use, Hostinger’s managed plan at $4.99 on a 24-month term with “AI credits included” in an amount the page does not state, OpenClaw Launch at $6 with $1 or $10 of credits, ClawHosters at €19 with three free models, Everpod at $29 including $10, RapidClaw at $29 including $20, Agent37 Plus at $9.99, Ampere by credits. Aplite is in the second group and has its own section below.

The model line: what a turn costs

Every message to an OpenClaw agent is at least one model call, and the call is bigger than the message. OpenClaw’s context documentation lists what the model receives on each turn: a system prompt that is “rendered each run” with the tool list, the skills list and the workspace bootstrap files (AGENTS.md, SOUL.md, IDENTITY.md, USER.md, up to 20,000 characters each and 60,000 in total by default), the tool schemas as JSON, which “count toward context even though you don’t see them as plain text”, the conversation history, every tool call and result, and “provider wrappers or hidden headers (not visible, still counted)”. The example on that page puts the system prompt at “38,412 chars (~9,603 tok)” and the tool schemas at “31,988 chars (~7,997 tok)” before a single word of conversation.

Our own fleet agrees. Across Aplite workspaces over the sixty days to 20 October 2026, on OpenClaw 2026.9.8 with a default workspace, the median chat turn sent about 16,000 input tokens and received about 250; the mean was 16,700 and 500. Scheduled runs were about 14,000 in and 210 out. Roughly half of the input tokens were cache reads, which the providers bill at a tenth of the input price or less. So the round figure for the arithmetic below is 15,000 input tokens and 400 output tokens per message, with caching ignored, which makes every number an upper bound.

Three models from Aplite’s catalog, priced from OpenRouter’s model list on 20 October 2026, per million tokens. Claude Haiku 5.5 is the cheap tier. DeepSeek V4 Flash 0731 is the platform default; OpenRouter routes it to 27 endpoints, most at $0.13 in and $0.26 to $0.40 out, the headline figure at $0.44 and $1.32, so it is shown as a range. Claude Sonnet 5.5 is a frontier model.

Claude Opus 5.5 at $4 and $20 doubles the Sonnet line: 6.8¢ a message, $61 a month. The :free variants on OpenRouter cost $0 and are rate-limited per account to 20 requests a minute and 50 a day, or 1,000 a day once you have bought $10 of credits; one agent turn with tools is several requests, so 50 a day is a short morning. Kilo’s article reports one operator’s three-week test of the same workload on different models: $8.70 a day on Opus 4.7, $2.80 on Sonnet 4.6, $0.12 on MiniMax M2.5. The spread is the model, not the server.

Two things change the per-message figure and both are in your hands. Longer sessions cost more, because the history rides along until compaction; and workspace files cost on every turn, which is why the docs cap them. The Hostinger guide’s observation that model choice matters more than server size is the right conclusion from the table above.

The hidden lines

The heartbeat. By default OpenClaw wakes the agent every 30 minutes to check whether anything needs attention, 48 times a day, and the heartbeat documentation is explicit: “Heartbeats run full agent turns. Shorter intervals burn more tokens.” It also gives the size: a run in the main session carries the conversation history, about 100K tokens, while isolatedSession: true and lightContext: true bring it down to “~2-5K per run”. On our fleet the isolated, light heartbeat has a median of about 2,200 input tokens. Output is a couple of hundred tokens and is ignored below.

On a frontier model the default heartbeat costs nearly ten times the thirty messages a day you actually send. This is the line behind the stories of an idle agent costing money: Kilo quotes a user billed “$5 a day for doing absolutely nothing” on an idle GPT-5.4 agent. The docs’ remedies are the config keys above, a cheaper model override for heartbeat runs (ollama/llama3.2:1b is their example), a longer interval, or 0m. Aplite’s runtime runs the main agent’s heartbeat every 4 hours, isolated and light, for every workspace, which is why the fleet figure is 2,200 and not 100,000 and six runs a day rather than 48.

Scheduled runs and sub-agents. A cron job is a full turn too, at the scheduled-run size above; a daily digest is one, an hourly watcher is 24. The FAQ on models notes that sub-agents parallelize work and “each sub-agent consumes its own tokens”. The Hostinger guide’s estimate that idle or forgotten automations are 10 to 30 % of spend is plausible against these sizes.

Workspace files and memory. The bootstrap files are injected on every turn up to the 60,000-character cap, about 15,000 tokens if you fill it, on top of the system prompt. A memory file that grows all month grows the bill all month.

Seeing the line. Self-hosted, /usage full adds an estimated cost to each reply and /usage cost prints today and the last 30 days; the usage tracking page notes that “sessions billed through a plan hide per-token dollar estimates”, so a subscription login shows tokens, not money. On Aplite the token usage shows under each reply and the money in Settings → Cost.

Your time. The line no provider prints. OpenClaw’s releases feed shows three stable versions in five days, 2026.9.33 on 15 October, 2026.10.4 on 16 October and 2026.10.5 on 19 October 2026; self-hosting means deciding about each, plus the first install, the systemd unit, the backups and the day the gateway stops answering. How much that is worth an hour is the one number on this page only you can fill in.

Where Aplite sits

Aplite runs a standard OpenClaw instance per workspace in Nuremberg and bills the models in credits at 1 credit = 1¢, spent only when an agent calls a model; the credit price of a model is above OpenRouter’s list price, which is how the hosting is paid for on Free. The Free plan is $0 with no card: one agent, 300 welcome credits that do not expire, and the budget end of the catalog, which on 20 October means Claude Haiku, Gemini 3.1 Flash Lite, Grok 4.3, DeepSeek V4 Flash and Qwen3.6 35B. The workspace sleeps after 6 hours idle and wakes on the next visit, and Free has no schedules and no credit packs, so when the 300 credits are gone the next step is Pro. Pro is $20 a month with 2,000 credits that reset monthly, six agents, always-on, schedules and the full catalog; packs run from $5 for 500 credits and never expire. Power is $50 with 6,000.

What that buys in messages, from the fleet figures above: a chat turn on the default model has cost 0.4 to 0.5 credits on average over the last sixty days, so the welcome grant is roughly 600 to 750 turns and Pro’s monthly credits roughly 4,000 to 5,000, before any scheduled runs; a Sonnet pin multiplies the per-turn cost by five to fifteen. What real workspaces spent, over the same sixty days: the median active day used about 5 credits and the busiest tenth of days about 22. Against the table, Pro at $20 is more than a $6 VPS with a cheap model and less than the same VPS with a frontier model and a default heartbeat; Free is $0 for as long as the credits last, with a gateway that is asleep when nobody is talking to it. Aplite is run by PIKsolution OÜ, an Estonian company, and is not affiliated with the OpenClaw project.

Prices on this page move. Every line carries the date it was read; the plan is to re-read them once a quarter.

FAQ

Is OpenClaw free?

The software is: OpenClaw is open source under the MIT license, and its site says “No subscription. No hosted tier. No token.” What costs money is the machine it runs on and the model it calls. The only setup with both at $0 is your own computer running a local model through Ollama, which the upstream FAQ warns is “usually no” for a good fit because small models truncate context; every other path pays for the server, the tokens or both.

How much do the models cost per month?

It depends on the model and on the heartbeat more than on how much you type. At thirty messages a day, with each message sending about 15,000 tokens of context, a month on Claude Haiku 5.5 is about $1.50, on DeepSeek V4 Flash $2 to $6.50, on Claude Sonnet 5.5 about $31, on Claude Opus 5.5 about $61, at OpenRouter’s prices on 20 October 2026. The default heartbeat, every 30 minutes in the main session at about 100K tokens a run, adds $14 a month on Haiku and $288 on Sonnet unless it is isolated or switched off.

What is the cheapest way to run OpenClaw 24/7?

In money, your own always-on machine with a local model: $0 plus electricity, which is under $1 a month for a Mac mini at idle. If the machine must be someone else’s, Oracle’s Always Free tier is $0 for a 2-OCPU, 12 GB Arm instance with a card at sign-up, and reclaims instances idle for seven days; the cheapest paid servers are $5 to $6.50 a month at Contabo, DigitalOcean and Hostinger, with renewals up to $11.99. The model tokens come on top in every case, from about $1.50 a month on a cheap model. Managed hosts with a free plan, Aplite among them, are $0 for the hosting with a credit allowance for the models; which of them keep a free plan and which take a card is in OpenClaw hosting with a free plan and no credit card.

How many tokens does one OpenClaw message use?

Far more than the message. OpenClaw’s context documentation shows a system prompt of about 9,600 tokens and tool schemas of about 8,000 before any conversation, and on Aplite’s fleet the median chat turn in October 2026 sent about 16,000 input tokens and received about 250. Half of the input is typically served from cache at a fraction of the price. A good planning figure is 15,000 in and 400 out per message.

Why is my OpenClaw bill so high?

Usually the heartbeat: by default the agent runs a full turn every 30 minutes, 48 a day, and in the main session each one carries the history, about 100K tokens. On a frontier model that is several dollars a day with nobody talking to the agent. Set isolatedSession: true and lightContext: true, give heartbeat runs a cheap model, or disable them with every: "0m". After that, look at scheduled jobs and sub-agents, which are full turns too, and at the size of the workspace files injected on every turn.

How much does Aplite cost per month?

Free is $0 with no card: one agent, 300 welcome credits that do not expire, budget models, and a workspace that sleeps after 6 hours idle. Pro is $20 a month with 2,000 credits, Power $50 with 6,000; 1 credit is 1¢ of model usage, priced above OpenRouter’s list. On the default model a chat turn has averaged 0.4 to 0.5 credits across our workspaces, so Pro’s credits are several thousand turns a month before scheduled runs.