OpenClaw Costs in 2026: What You Actually Pay

Do not index

OpenClaw costs come in three lines: the software, which is free and open source; the model API tokens your agent burns; and an always-on computer to run it. For most people the honest 2026 math is $0 for OpenClaw itself, a usage-driven token bill, and about $6 to $25 a month for hosting, or from $3.99 managed. We run Agent 37, a managed OpenClaw host, so we watch real cost data across more than 1,000 live instances. This guide breaks down every line, ours included.

How much does OpenClaw cost per month?

There is no OpenClaw subscription. Anyone selling you "OpenClaw pricing plans" is selling hosting. The full bill at a glance:

OpenClaw software: $0. Free, open source, no license fee, no trial clock.

Model API usage: variable. You bring your own key (Anthropic, OpenAI, Google, or a router) and your provider bills exactly what your agent uses. Light personal use costs pocket change. Heavy automation is real money, and it is usually the biggest line.

Always-on compute: $4 to $25 for most setups. From $3.44 a month metered on Agent 37 Cloud, $3.99 managed, roughly $6 to $25 on a DIY VPS, and $25 to $33 for the same box on AWS or GCP.

Setup time. One click on a managed platform. An evening to a weekend on a VPS, plus the ongoing habit of keeping it alive.

If a quiet month costs you more than about $10 in hosting, you are overpaying for the compute. If a busy month surprises you, the surprise is almost always tokens. The rest of this post prices both lines properly.

What makes up OpenClaw costs?

Think in three layers:

Model API usage. Every reply, tool call, and long context window burns tokens. Costs scale with how hard your agent works, not with time. This is the line you control with model choice and usage caps.

Always-on hosting. OpenClaw is a long-running process. It has to be awake when a WhatsApp message lands at 3 a.m. That rules out your laptop and rules out serverless platforms that sleep on idle. You need a persistent Linux environment running 24/7.

Ops time. Updates, TLS, out-of-memory recoveries, and restarts. DIY hides this line until the weekend it stops being free.

For the official metering model, see the OpenClaw API usage and costs docs. Two line items people miss: built-in web search is billed per request once you pass the monthly free allowance, and scheduled heartbeats call your model on a timer even when nobody is messaging the agent.

Where do most OpenClaw costs go: API or hosting?

For an active agent, tokens beat the server. A $6 VPS can look cheap while the API bill runs several times higher. Switching your default model or tightening prompts often saves more than any hosting change.

Do not overbuy CPU. Inference happens at the model provider, not on your box.

Measure token spend for a week before you optimize hosting.

Treat hosting as the reliability line, not the savings lever.

We keep a verified, regularly refreshed price table of which models actually work well with OpenClaw in our best AI models for OpenClaw guide, including what a realistic month of tokens costs per model.

The hidden OpenClaw costs: token burn, not server bills

Every OpenClaw cost horror story on Reddit follows the same shape: the server bill was fine, the token bill was not. The usual causes:

Heartbeats on an expensive model. A heartbeat that pings a frontier model every few minutes quietly bills you around the clock. Point heartbeats and background jobs at a cheap model.

Runaway tool loops. An agent that retries a failing task all night can burn a month of hosting budget before breakfast. Cap retries and loop depth.

Oversized context. Long memory files and giant system prompts are re-sent on every call. Trim them and the same conversations cost a fraction.

Cron jobs you forgot. Scheduled tasks bill like usage because they are usage. Review what actually needs to run.

Set a spend alert at your model provider before your agent gets busy. It is one click, and it turns a $200 surprise into a $10 lesson.

What OpenClaw needs from a host (and why it changes the price)

OpenClaw's own FAQ keeps requirements modest: a small VPS or Raspberry Pi-class box is fine, and 4 GB of RAM is plenty. Our fleet data agrees, with one nuance. Measured across roughly 1,000 live instances, an idle instance sits near 250 MB of RAM, then spikes toward 4 GB when tools fire and a browser session opens. Instances spend about 90 percent of their time idle.

That spiky, mostly-idle profile breaks the assumptions cheap hosting is priced on:

Shared hosting throttles or kills processes during memory spikes. Your agent fails silently.

Serverless sleeps between requests. An agent that sleeps misses messages.

Standard VPS sizing makes you pay for peak RAM around the clock to cover spikes that happen a few minutes a day.

So the checklist for any OpenClaw host: a persistent process, writable disk for configs and skills, 4 GB of RAM available if browser automation is in scope, stable networking, and no silent throttling. Cheap hosting that fails any of these is expensive hosting with a small invoice.

OpenClaw hosting costs compared: every option in 2026

Option	Typical monthly price	Setup	The catch
Old laptop, Mac mini, or Pi	$0 plus electricity	An afternoon	Home power and network outages become agent outages
Shared hosting	$5 to $15	Minutes	No guaranteed RAM; background processes get killed
Free cloud tier (Oracle)	$0 sticker	Hours	Now capped at 2 OCPU / 12 GB; idle instances get reclaimed
DIY VPS (Hetzner, Hostinger, DigitalOcean)	$6 to $25	An evening to a weekend	You own SSL, updates, monitoring, and 3 a.m. incidents
Big-cloud VM (AWS, GCP)	$25 to $33	Hours	5x to 7x a specialist host for the same always-on box
Agent 37 managed	$3.99 to $99.99	One click	No free tier; support is AI-filtered on cheap tiers
Agent 37 Cloud (API)	From $3.44, metered per minute	One API call	Prepaid balance; you drive it via API

The 2026 price moves matter here. Hetzner raised cloud prices on June 15, 2026: the cheapest plan is now about 5.49 euros in EU locations, and US locations start near $23. Hostinger's entry VPS is $6.49 a month on a two-year prepay and renews at $11.99. Oracle quietly halved its always-free ARM tier to 2 OCPU and 12 GB. And an always-on 4 GB box runs about $25 on GCP and $30 or more on AWS before storage. Platforms built for ephemeral code sandboxes are the wrong tool entirely: always-on, they price around $121 a month. Full DIY tradeoffs live in our VPS for OpenClaw guide and the Hostinger OpenClaw teardown.

Two patterns jump out of the table. General-purpose clouds charge a heavy premium for always-on workloads. And the cheapest cash price (a bare VPS) and the cheapest total price (managed) are different answers, depending entirely on whether you enjoy operating Linux servers.

How much is Agent 37 for OpenClaw?

Full disclosure: Agent 37 is our product, and more than 1,000 people pay for it today. These are the live prices, tradeoffs included:

Tier	Monthly price	What you get
Basic	$3.99	1 vCPU, 4 GB RAM, 12 GB disk, bring your own API keys, $1/mo AI credits, 75 web searches and 1,000 app actions monthly, community support
Plus	$9.99	Adds bundled free models (GPT-OSS 20B and 120B, DeepSeek V4 Flash) capped per rolling 5-hour window inside a weekly cap, 20 GB disk
Pro	$29.99	Adds a built-in Chromium browser for web automation, 2 vCPU, 8 GB RAM, 30 GB disk, $5/mo credits, email and chat support
Max	$99.99	4 vCPU, 16 GB RAM, 60 GB disk, $20/mo credits, team roles, priority support, dedicated onboarding

Every plan runs your agent in an isolated container managed from the browser: task board, full web terminal, visual file browser, a live Linux desktop, scheduled jobs, and 1,000+ app integrations through Composio. BYOK means API keys go straight from your container to the model provider; we never sit in the middle of your token bill. Runtime updates and security patches roll out automatically, and you can cancel anytime.

Two things to know before you buy. There is no free tier; $3.99 a month is the cheapest way in. And the honest tradeoff at these margins is support: an AI bot filters requests first on the low tiers. Current plan details live on the pricing page. For the wider market, see our comparison of OpenClaw hosting providers.

If you run a fleet, the unit economics compound. Ten instances at $3.99 cost $39.90 a month. Ten mid-range VPSes at $15 cost $150. That difference is your margin.

How we host OpenClaw for under 99 cents

This is the engineering behind those prices, and it is repeatable if you like building infrastructure. We set a target that sounded absurd: get the hosting cost of a real, always-on OpenClaw instance under 99 cents a month. Real execution, real uptime, real users.

The unlock was the workload data above. If an idle instance uses 250 MB and a busy one needs 4 GB, you should not buy one small VM per instance. You do the opposite: buy the cheapest large-memory machines you can justify (we used 32 GB boxes) and pack many isolated containers onto each.

Pack on average behavior, not worst case. Not every instance spikes at once, so oversubscribe RAM on measured averages, the way real infrastructure providers do.

Cap every container at 1 vCPU and up to 4 GB RAM. Without hard caps, one runaway workload takes down every neighbor on the node.

Shift load at peak. Lightweight orchestration moves containers to another machine when a node runs hot, so you never pay for peak capacity around the clock.

Treat restarts as routine. Persistent disks and fast recreation keep data intact and downtime to seconds.

The result: a measured per-instance hosting cost under 99 cents a month in real usage conditions. The same architecture, bare metal plus disciplined oversubscription, is why a managed plan can sell at $3.99 with reserved resources while a general-purpose cloud charges $25 for the same job. It was not magic. It was infrastructure math. The raw version of this story, messy details included, is in our Reddit thread.

How can you cut OpenClaw costs without killing quality?

Pick a cheaper default model for routine traffic and keep frontier models for hard steps. Model routing is the single biggest lever.

Point heartbeats and scheduled jobs at your cheapest model.

Cap tool loops and trim long context. Both bill on every call.

Right-size hosting. A quiet agent does not need a 16 GB box.

Use free tiers for labs, not production. Reclaim risk is a cost too. Our free OpenClaw hosting guide compares the genuinely free routes honestly.

Prefer managed when your hourly rate makes DIY debugging more expensive than the plan.

Before committing to any host, get written answers to four questions: how long a task can run before termination, what happens during a usage spike, whether background processes pause during quiet hours, and how failures are surfaced. Be suspicious of "unlimited". Providers that publish explicit limits are almost always cheaper in practice.

OpenClaw costs FAQ

Is OpenClaw free?

The software is free and open source. Your money goes to model API usage and always-on hosting. We wrote a full honest answer to "is OpenClaw free" if that is your actual question.

What is a realistic monthly OpenClaw budget?

A quiet setup: $4 to $10 of hosting and a few dollars of tokens. An active personal agent: $10 to $50 all-in, with tokens dominating. Heavy automation: the token line grows without a ceiling, which is why spend alerts matter more than hosting choice.

Do I pay Agent 37 for model tokens?

No. You bring your own API keys, and your model provider bills token usage directly. Agent 37 bills only the managed hosting. Plus tiers and above also bundle capped free models if you want a $0 token line for background work.

How much RAM does OpenClaw need?

OpenClaw's FAQ says 4 GB is plenty, and our fleet numbers back that: around 250 MB idle, spiking toward 4 GB under load with browser automation. A 2 GB box works until the first heavy browser session, then falls over.

Is a free VPS the cheapest OpenClaw path?

On sticker price, yes. On total cost, often not. Oracle's always-free ARM tier was cut to 2 OCPU and 12 GB in mid-2026, and over-limit free instances face reclamation. Free is fine for experiments; it is a bad place for an agent you rely on.

Why is big-cloud hosting so much more expensive?

AWS and GCP price for general-purpose workloads: an always-on 4 GB VM runs about $25 to $33 a month with storage, roughly 5x to 7x what a specialist host charges for the same job. You are paying for flexibility your agent never uses.

The bottom line on OpenClaw costs

OpenClaw is free. Pay your model provider for what your agent actually uses, put a spend alert on that line, and pay as little as possible for a box that never sleeps. If you want the always-on part solved in one click, Agent 37 runs your OpenClaw instance from $3.99 a month: isolated container, web terminal, live desktop, cancel anytime. Building agents into your own product instead? Agent 37 Cloud gives you the same always-on instances behind one API from $3.44 a month, metered per minute, with a $1 starter credit and no card needed to try it.

Prices are half the story. The complete OpenClaw hosting guide covers setup, security, and scaling for every option here.