How much does an always-on AI agent cost?

An always-on AI agent costs the sum of its host, model usage, paid tools, storage and backups, and operating time. Keeping the process available does not mean it spends tokens continuously. The bill depends on what wakes it, how much context it sends, and how many attempts it makes before finishing.

The worked example below comes to $72.60 per month in infrastructure and model charges, or $222.60 with an illustrative labor allowance. These are calculated scenarios, not a cloudrun quote, customer average, or measured benchmark. Every workload and labor assumption is stated so you can replace it.

Separate fixed costs from usage

Start with this equation:

Monthly operating cost =
  hosting + backups + extra storage + paid tools
  + model usage + operating labor

Use a separate line for initial setup. Spreading setup across a year can help compare options, but it should remain visible rather than disappearing into the monthly total.

Keep cash spending and labor valuation separate. A founder doing maintenance personally may pay no additional wage that month, but the time still competes with client work or product development.

The example's prices and assumptions

Prices were checked on September 5, 2026. The host reference is DigitalOcean's regular Basic Droplet with 4 GiB memory, 2 vCPUs, and an 80 GiB SSD, listed at $24 per month. This is a budgeting example, not evidence that the machine meets a particular agent's requirements. DigitalOcean Droplet pricing

The backup reference uses DigitalOcean's basic daily plan, priced at 30% of the Droplet's monthly cost. In this example, $24 × 30% = $7.20. A backup price does not establish application consistency or successful restoration. DigitalOcean backup pricing

For model usage, the example uses Claude Sonnet 4.6 at the listed base API rates of $3 per million input tokens and $15 per million output tokens. It assumes ordinary uncached requests and excludes cache writes, cache reads, batch discounts, premium modes, and separately charged tools. The same live pricing page lists other models and rates; Sonnet 4.6 is simply the named model used for reproducible arithmetic here. Anthropic API pricing

All amounts are USD, before taxes. The workload period is 30 days. Included transfer and disk capacity are assumed sufficient; extra storage, transfer overages, paid tools, and setup fees are set to zero for this example. Replace those zeroes when your workflow uses them.

Calculate model usage across the whole task

Count all model calls needed to finish a task, including retries and tool-result follow-ups. A user's single message is not a reliable billing unit.

Model cost =
  (uncached input tokens / 1,000,000 × input rate)
  + (output tokens / 1,000,000 × output rate)

If you use caching, add the provider's separately reported cache-read and cache-write quantities at their applicable rates. Do not charge the same token quantity twice by treating already separated usage fields as additional uncached input.

For an illustrative modest workload, assume 20 completed tasks per day. Each task consumes 10,000 input tokens and 2,000 output tokens in total across all its calls. These are invented planning inputs, not observed averages.

Tasks per month: 20 × 30 = 600
Input tokens: 600 × 10,000 = 6,000,000
Output tokens: 600 × 2,000 = 1,200,000
Model cost: (6 × $3) + (1.2 × $15) = $36

Now assume a model-based scheduled check runs 24 times per day, using 2,000 input tokens and 100 output tokens each time:

Checks: 24 × 30 = 720
Input cost: 720 × 2,000 / 1,000,000 × $3 = $4.32
Output cost: 720 × 100 / 1,000,000 × $15 = $1.08
Scheduled-check model cost: $5.40
Total model cost: $36 + $5.40 = $41.40

If a conventional scheduler checks a timestamp without calling a model, omit this token line. If it launches additional agent work, count that work too.

Three workload scenarios on the same host

The table holds the host and daily-backup example at $31.20. Every scenario uses the same token rates and 30-day month. The busy scenario assumes 100 tasks per day at 30,000 input and 4,000 output tokens per task, plus the same scheduled checks.

For the busy case, 3,000 tasks consume 90 million input tokens and 12 million output tokens. That is $270 + $180 = $450, plus $5.40 for checks. Holding the host constant isolates the effect of usage; it does not show that this machine can handle the busy workload. Capacity must be tested separately.

The $75 hourly valuation and monthly hours are assumptions. Substitute your actual cost and time records. These totals are not minimums, maximums, or a claim about typical spending.

Compare the model with a managed deployment

Use the same workload assumptions when comparing cloudrun with a VPS: hosting is one line, model usage is another, and someone still needs to review outputs and maintain the agent's configuration. Compare the current cloudrun plan with the complete operating budget above.

For a client service, record who approves spending and who responds to a failed run in the client-agent launch checklist. The agency setup guide gives you a starting point for evaluating the workspace alongside that service budget.

When you have a budget and a test workload, create an account to evaluate the setup. Judge the result by the work completed and the total cost of operating it.

Why a small change can alter the bill

Suppose the modest workload keeps the same task count and output, but the input per task grows from 10,000 to 40,000 tokens. Task input rises from 6 million to 24 million tokens. At the example rate, that adds $54 per month before any change in output or retries.

A new document attachment, longer conversation history, or extra tool result can change the amount sent to the model. Inspect request usage rather than estimating cost from the length of the final answer.

Retries deserve their own measurement. Record cost per accepted result alongside cost per attempt. A cheaper individual call can still be expensive if it repeatedly fails the workflow's acceptance criteria.

Compare managed hosting with the same accounting

Replace the VPS and backup lines with the managed plan's actual quote, then check what it includes. Confirm whether model usage is bundled, metered separately, or billed directly through your own API account. Add charges for seats, agents, disk, or tools where the plan applies them.

Use the hosting responsibility checklist to check which maintenance and recovery tasks each quote covers.

For an illustrative comparison, suppose a managed plan costs $60 per month, including backups equivalent to your requirement. Its premium over the $31.20 infrastructure example is $28.80. At an assumed $75 per hour, it needs to save about 23 minutes of operating time each month to offset that premium. This is a hypothetical quote and a break-even calculation, not a measured time saving.

Build your first real budget

Run a representative sample for a week. Keep model and tool usage associated with the workflow, record failed attempts, and note maintenance time. Project the usage to a month, then separately test a busier scenario rather than assuming the sample captured every future workload.

Set budget alerts and decide which work pauses when a threshold is reached. Test the pause: an email warning alone does not stop a loop. Review the estimate after changing the model, adding a tool, or expanding the agent's schedule.

For several client agents, record each budget and stop rule in the fleet register described in the agency playbook.

Does bringing my own API key make model usage free?

No. It changes which account receives the model charges. Enter the provider's invoice in the same cost model and confirm whether the hosting service adds any separate usage fee.

Can an agent cost money while nobody is chatting?

Yes, if it runs scheduled model calls or other billable work. An idle rented server can also have a hosting charge. Inspect the configured schedules and the provider's usage records to distinguish the two.