How Much Does OpenClaw Cost to Run? (2026 Breakdown)

OpenClaw is free software running on a $4 server, and people still open bills for $85 a month. The gap is never the hosting. It’s the model API — and specifically two settings almost nobody checks before deploying: how often the agent wakes itself up, and whether its system prompt is being cached.

The short version:

• The server is $3–12/month and is the smaller half of the bill for almost everyone.
• Model tokens are $2–15/month for genuine personal use — and can run to hundreds if you pick a frontier model and leave the defaults alone.
• The heartbeat is the classic trap. A background check-in every 30 minutes costs nothing on a cheap model and roughly $85/month on a frontier one, before you run a single task.
• Prompt caching is the biggest lever nobody pulls. An agent re-sends the same system prompt on every call; cached reads cost about a tenth of normal input tokens.

The infrastructure lines this article doesn’t cover — backups, snapshots that outlive the server, IPs that bill when idle, and the hardening hours — are itemised in the hidden costs of self-hosting an AI agent.

Wondering whether to hand the whole thing to someone else? Managed OpenClaw vs self-hosted runs the same arithmetic against managed plans — and the answer flips on the token bill below, not on the server.

The four things you actually pay for

Everything below is the detail behind those four rows.

The server: $3–12/month

This part is easy and cheap, and the market is competitive enough that the differences are small in absolute terms.

Contabo is the RAM-per-dollar leader — its dedicated OpenClaw tiers start around $4.50/month and it consistently ships several times the memory per dollar of the premium hosts. The trade is shared CPU that can throttle under sustained load, which matters more for an always-on agent than for a website.

Hostinger is the one to pick if you want to skip the setup entirely: its VPS Docker Manager ships a one-click OpenClaw template, and the entry plan is 8GB of RAM at $6.49/month — correctly sized for browser automation rather than needing an upgrade a fortnight in. It also bundles inference credits, which is convenient to start with and carries a margin over buying tokens direct.

Hetzner is the community default at roughly $3.49–4.49/month for 2 vCPU and 4GB, with hourly billing and a monthly spend cap. No one-click install, so budget 40 minutes.

Elestio sells managed OpenClaw hosting — someone else owns patching, updates and uptime. You pay a multiple of the raw VPS price for that, which is rational if your time is worth more than the difference.

Oracle Cloud is genuinely free forever: the Ampere allocation is 2 OCPUs and 12GB of RAM with 200GB of storage, per Oracle’s documentation — not the 4 OCPU / 24GB figure most articles repeat. On paper it’s the best machine here. In practice free capacity is frequently unavailable, the console is built for enterprise infrastructure teams, and you are locked to one region.

If you’re using Hostinger’s bundled nexos.ai credits rather than your own API key, there’s a further wrinkle: Hostinger publishes how many credits each model consumes but not what a credit costs, so the markup isn’t computable from the docs. We work through what can and can’t be established in the OpenClaw template cost breakdown.

The honest advice: the spread between the cheapest and most expensive option here is about eight dollars a month. Do not spend an evening optimising it. Buy 8GB, get on with it, and spend the attention on the section below — that’s where the money is.

The model API: where the bill actually comes from

OpenClaw ships no intelligence. Every message it processes, every tool call, every document it reads is billed by whichever model provider you connect. Pricing spans more than two orders of magnitude across the market, so this single choice dominates everything else.

Current published rates for Anthropic’s lineup, per million tokens:

Budget models from other providers run considerably cheaper — sub-$0.10 per million input tokens at the very low end — and frontier tiers run higher. The 100× spread between the cheapest and most expensive option is the reason two people running “the same” agent can report wildly different bills.

What a task costs. A medium-complexity agent task typically takes 3–8 model calls and consumes somewhere around 80,000–150,000 input tokens once you count the system prompt, conversation history and tool results being re-sent each time. Ten of those a day is roughly 30–45 million input tokens a month.

Run that on Haiku at $1 per million and you’re looking at tens of dollars. Run it on a frontier model at $5 per million and the same workload is several times that. Nothing about your agent changed — only the model string.

The heartbeat: the single most common surprise

This one catches almost everybody, and it is worth checking before you deploy anything.

OpenClaw’s default configuration includes a heartbeat that periodically wakes the agent to check in. Out of the box it fires every 30 minutes and consumes roughly 8,000–15,000 tokens per call, because it carries context with it. Over 24 hours that is around 570,000 input tokens before your agent does a single useful thing.

On a budget model that’s pennies. On a frontier model it is roughly $2.85 a day — about $85 a month to keep an idle agent checking its watch.

Three ways to deal with it, in order of how much they save:

1. Turn it off. Set the interval to zero. For a personal assistant you invoke on demand, background wake-ups buy you nothing.
2. Slim it down. Isolated sessions with light context cut each heartbeat to roughly 2,000–5,000 tokens — about a 90% reduction — while keeping proactive behaviour.
3. Point it at a cheap model. If you want the agent genuinely proactive, the check-in does not need your best model. Reserve that for real work.

Prompt caching: the lever most cost guides miss

Here is the thing that almost no OpenClaw cost breakdown mentions, and it is arguably the biggest one.

An agent sends its system prompt, its tool definitions and its accumulated conversation history on every single call. That prefix is largely identical each time. Without caching you pay full input price for all of it, every time.

With prompt caching enabled, that repeated prefix is written to a cache once and read back at roughly a tenth of the normal input price. The write costs about 1.25× normal input rates, so with a five-minute cache window you break even at two requests and save on everything after. For an agent making 3–8 calls per task, and many tasks per day, against a system prompt that never changes, this is not a marginal optimisation — the repeated prefix is usually the majority of your input tokens.

Two practical notes. Caching is a prefix match, so anything that changes at the front of the prompt invalidates everything after it — the classic mistake is interpolating the current timestamp or a session ID into the system prompt, which silently makes the whole thing uncacheable. And there is a minimum cacheable length (roughly 512–1,024 tokens depending on the model), below which nothing is cached and no error is raised.

If your bill looks higher than the arithmetic above suggests, check caching before you switch models.

Four realistic scenarios

Personal use — $7 to $15/month. A Hetzner or Contabo box at $3.50–4.50, a budget or mid-tier model, 5–10 tasks a day, heartbeat off. LLM spend $2–10. This covers email drafting, research summaries and light automation, and it is where most individuals land.

Power user — $25 to $80/month. A slightly larger VPS at $4.50–9, Haiku-class or comparable, 20–50 tasks a day, a slimmed-down heartbeat. LLM spend $20–75. A developer or freelancer running coding assistance and document analysis through the working day.

Small team — $60 to $200/month. A 48GB box around $25, a Sonnet-class model, several users, 50–200 tasks a day. LLM spend $35–175. The memory matters here because concurrent agent sessions must not swap.

High volume — $500 to $1,500+/month. Dedicated hardware or a large instance, frontier models or mixed routing, 500+ tasks a day. LLM spend is essentially all of it. At this scale model routing stops being optional.

Notice the pattern: hosting is a rounding error in every tier except the first, and by the top tier it is under 5% of the total.

Cutting the bill, in order of impact

1. Route by task complexity. Not everything needs a frontier model. Simple lookups, formatting and templating run fine on a cheap one; reserve the expensive model for multi-step reasoning and code. Teams that implement routing typically cut LLM spend by 40–60%, and it is the single largest saving available.
2. Enable prompt caching and keep the front of your prompt static. See above — for a repetitive agent workload this can be transformative, and it costs nothing to turn on.
3. Kill or shrink the heartbeat. Potentially $85/month on a frontier model for zero work delivered.
4. Right-size the server. If your agent runs ten tasks a day it does not need 48GB. Start at the smallest plan that meets your actual concurrency and upgrade on evidence, not anticipation.
5. Measure before committing. Track real token consumption for a week before locking in a model tier. Most people overestimate their needs and pay frontier prices for work a mid-tier model handles identically.

What about running a local model?

The reasoning is appealing — eliminate per-token fees entirely — and it is usually wrong at personal scale.

A local model swaps a variable bill for a fixed one. You need 8GB of RAM for a 7–8B parameter model and 16GB or more for a 13B, and on CPU alone generation is slow enough that most people give up. Making it genuinely usable means a GPU, which is $150–500/month for a dedicated instance or roughly $0.15–0.80/hour on demand.

Set that against $2–15/month in API calls and the arithmetic is stark. Local inference makes sense for three reasons: data that legally cannot leave your infrastructure, offline operation, or volume high enough that per-token fees exceed the hardware. It does not make sense as a way to save money on a personal assistant.

The bottom line

For a personal agent, budget $10–25/month all in — a $6 server and $5–15 of tokens, assuming you turned the heartbeat off and left caching on. For serious daily work with a good model, $30–80. Local inference starts at $200+ and is a compliance decision rather than a cost one.

The decision that matters is not which host to buy. It’s which model you point the agent at, whether it wakes itself up for no reason, and whether you’re paying full price for the same system prompt several thousand times a month.

FAQ

Is OpenClaw free?

The software is free and open source with no paid tier or subscription. You pay for the infrastructure around it: a server if you host it remotely ($3–12/month) and API tokens for whichever language model you connect ($2–15/month for typical personal use). Running it on hardware you already own reduces the cost to just the tokens.

Why is my OpenClaw bill so high?

In order of likelihood: the heartbeat is still on and hitting an expensive model (up to ~$85/month for no useful work), prompt caching is off so you’re paying full price to re-send the same system prompt on every call, or you’re routing every task to a frontier model when most of them would run identically on a cheaper one. Check those three before changing anything else — the server is almost never the problem.

How much does the OpenClaw server cost?

Between $3 and $12 a month for a suitable VPS. Around $3.50 gets 2 vCPU and 4GB, which is enough for a headless agent; $6–9 gets you 8GB, which is what you need once browser automation is involved. Oracle Cloud’s free tier can run it for nothing if you can get capacity and tolerate the complexity.

Is it cheaper to run a local model instead of paying for API calls?

Rarely, at personal scale. A local model removes per-token fees but requires 8GB of RAM for a small model, 16GB+ for a mid-size one, and realistically a GPU at $150–500/month to be usable. Against $2–15/month in API costs, the crossover only arrives at sustained high volume — or when data residency requirements mean cost was never the deciding factor.

Related Blog Posts

• The Best VPS for OpenClaw
• What Is OpenClaw?
• How to Self-Host OpenClaw on a VPS
• The Best VPS for AI Agents