"Can I run this agent around the clock, and what will it cost me?" is one of the most common questions buyers ask, and most pricing pages answer it badly. They quote a monthly plan fee, which tells you almost nothing, because two agents on the same plan doing the same job can differ in cost by two orders of magnitude.
The number that matters is how often your agent actually executes. This post works through the arithmetic.
What "running 24/7" actually means
There is a common assumption that an always-on agent is like an always-on server: something is running, burning resources, every second of the month. For almost every agent platform, that is not how it works.
An agent at rest costs nothing to keep available. It is a stored configuration waiting for a trigger. The cost begins when it executes. So "running 24/7" really means "available to execute at any hour", and your bill is set by how many executions happen in those hours, not by the hours themselves.
There is one important exception. If you self-host, you are paying for a virtual server by the hour whether or not your agent ever fires. That is a genuine fixed 24/7 cost, and it is why a self-hosted agent that runs twice a day is usually more expensive than a managed one. We cover that trade in the cheapest AI agent platforms.
The arithmetic: how often does your agent really run?
Start with executions per month, because every other number follows from it. A month is roughly 30 days, or 720 hours.
| Trigger pattern | Runs per day | Runs per month | At 1 cent per run | At 5 cents per run |
|---|---|---|---|---|
| Poll every minute | 1,440 | 43,200 | $432 | $2,160 |
| Poll every 5 minutes | 288 | 8,640 | $86 | $432 |
| Poll every 15 minutes | 96 | 2,880 | $29 | $144 |
| Poll hourly | 24 | 720 | $7 | $36 |
| Event-triggered (~20 real events/day) | 20 | 600 | $6 | $30 |
| Scheduled once daily | 1 | 30 | $0.30 | $1.50 |
The per-run range of roughly 1 to 10 cents covers a typical recurring agent run on frontier-model rates. Cheaper models and shorter prompts land below it; long multi-step reasoning runs land above it.
Read the first and fifth rows together. Both describe an agent that responds to about twenty meaningful events a day. One costs $432 a month and one costs $6. The work delivered is identical. The entire difference is that the first agent asks "has anything changed?" 43,200 times and the second is told when something changed.
Why most polls are wasted
Consider an agent watching an inbox for invoices. A busy shared inbox might receive 30 relevant emails a day. If the agent polls every minute, it wakes 1,440 times a day to find something 30 times. About 98% of its executions find nothing at all.
This is the single most common way an agent bill surprises someone. The agent works correctly. It is simply asking a question that is almost always answered "no", and paying full price for each answer.
The fix is not to poll less often, which just makes the agent slower to react. The fix is to make the "no" answer cheap.
Gate the model behind a cheap check
The expensive part of an agent run is the language model call, not the wake-up. A well-built recurring agent separates the two:
- Cheap check first. Query the API for new items since the last watermark. This is an ordinary API request costing a fraction of a cent, or nothing.
- Exit early if nothing changed. No model call, no meaningful cost.
- Invoke the model only on real work. The full reasoning run happens 30 times a day instead of 1,440.
With that structure, a minute-level polling agent costs close to the event-triggered row of the table above while keeping minute-level responsiveness. You are paying for the work, not for the watching.
Better still, use a real event trigger where the tool offers one. Most major SaaS tools can push a webhook when something changes, which removes the polling question entirely.
The other multiplier: retries
Retries deserve their own line in any 24/7 cost estimate because they break the arithmetic above. A failed run that automatically retries costs tokens each time. An agent that hits a persistently failing dependency at 2am and retries in a tight loop can execute thousands of times before anyone notices in the morning.
Three controls prevent this, and all three are worth setting before you leave an agent running unattended:
- A retry cap per run, with exponential backoff rather than immediate re-fire.
- A hard monthly spend ceiling that stops execution rather than just alerting.
- A daily spend check for the first month, which is when misjudged trigger frequency shows up.
More detail on controlling ongoing spend is in AI agent cost optimization, and on tracing spend back to specific agents in AI agent cost attribution.
What this looks like in practice
When we look at agents that people set up to run continuously on Gravity, the pattern is consistent: the ones that surprise their owners on cost are almost never the ones doing the most work. They are the ones checking most often for work that is not there.
The two recurring shapes are a monitoring agent polling a source that changes a few times a week, and a "watcher" duplicating something the source tool already offers as a notification. In both cases the fix is the same and takes minutes: widen the interval, or move to an event trigger, and put the cheap check in front of the model call. The agent keeps doing its job and the bill drops by an order of magnitude.
This is also why we price the way we do. Gravity's paid plans start at $20 per month with $20 of usage included, and extra usage can be added beyond the plan. A predictable plan fee means an always-on agent does not turn into an unpredictable bill, which is the actual anxiety behind the question.
How to estimate your own 24/7 cost
Four questions produce a usable number in a few minutes:
- How often does the trigger fire? Not how often you want a result, but how often the agent executes. Use the table above.
- How much of that is real work? Multiply real events by your per-run cost; treat empty checks separately, because they should be near-free if the agent is built well.
- How long is a run? A single-step summary and a ten-step research task differ by an order of magnitude in tokens.
- What is the failure mode? Assume one bad night of retries per quarter and confirm your cap would have caught it.
If the resulting number is uncomfortable, the lever is almost always question one. For how the underlying pricing models compare, see AI agent pricing in 2026, and for the equivalent question about replacing a role rather than a task, see how much an AI employee costs.
Frequently asked questions
How much does it cost to run an AI agent 24/7?
Typically $5 to $50 per month for an event-triggered or sensibly scheduled agent. A minute-level polling agent that calls the model on every wake-up can reach several hundred dollars a month doing the same work, which is why trigger design matters more than plan choice.
Do I pay while the agent is idle?
On a managed platform, no. An agent waiting for a trigger is a stored configuration, and billing starts at execution. On a self-hosted setup you pay for the server every hour regardless, which is the real fixed cost of always-on.
Is it cheaper to run an agent constantly or on a schedule?
A schedule is cheaper than constant polling, and an event trigger is cheaper than both. The ranking follows executions per month, so anything that reduces pointless executions reduces the bill proportionally.
How much does an AI agent cost per hour?
Per-hour framing is misleading, because an idle hour costs nothing on a managed platform. If you need an hourly figure, divide monthly executions by 720. An agent polling every 5 minutes performs 12 executions an hour, which is roughly 12 to 60 cents an hour if every one invokes the model, and far less if empty checks exit early.
Why did my always-on agent cost more than expected?
The two usual causes are polling frequency set higher than the work requires, and retries firing without a cap. Check executions per month first; if that number is much larger than the number of real events handled, the agent is paying to ask a question that is nearly always answered "no".
Can an agent run 24/7 on a free tier?
For light, low-frequency work, often yes. Free tiers usually cover one agent with a capped amount of use, which suits a daily or hourly job better than a minute-level one. Gravity's free tier is $0 per month for one agent.
Sources
- Execution counts are arithmetic from a 30 day month (720 hours). Per-run cost ranges reflect typical recurring agent runs at frontier-model rates.
- Current model rates: OpenAI pricing, Anthropic pricing.
- Related: AI agent pricing in 2026, cheapest AI agent platforms, AI agent cost optimization.