AI agent cost per month: tokens, hosting and review time
What one AI agent costs per month at 1,000 and 10,000 runs: API tokens on three models, prompt caching, hosting, monitoring, review time and upkeep.
In the example below, most of the AI agent cost per month is people’s time: reviewing the runs the agent flags and keeping it working.
In short:
- In the worked example below, one invoice intake agent costs about $830 to $1,830 a month at 1,000 runs, and about $3,180 to $4,350 at 10,000 runs. The range depends mostly on who looks after it.
- API tokens are a small line at that size: $18 to $36 a month for 1,000 runs on Claude Sonnet 5.5, GPT-6.1 Terra or Gemini 3.8 Flash, with prompt caching on.
- Human review is the biggest line that grows with volume. If a person checks one run in five for 3 minutes, that’s $244 a month at 1,000 runs and $2,436 at 10,000, at the US median wage for bookkeeping clerks.
- Prompt caching cut the token cost per run by 44 to 46 percent in this example.
- List prices change. Gemini 3.8 Flash doubles in price on 1 January 2027.
These figures are calculated from list prices and the assumptions stated in each section. None of them are measured. Replace the assumptions with numbers from your own logs before you budget.
What goes into the monthly bill for one AI agent?
Five lines. The model you pick mainly changes the first two:
- API tokens, priced per million tokens read (input) and written (output).
- Prompt caching, which lowers the price of input the model has already seen.
- Hosting for the code that runs the agent, its queue and its Slack or email connections.
- Monitoring, so you can see each run’s steps, costs and failures.
- People: someone reviews the runs the agent flags, and someone keeps the agent working as models and prices change.
How many tokens does one agent run use?
An agent run is several model calls in a loop, and every call sends the whole conversation again. Our example is an invoice intake agent of the kind described in when an invoice agent must ask a human. Each run reads one invoice email, looks up the supplier, checks for duplicates, then creates a draft bill or asks for approval.
Assumptions for one run:
| Part | Tokens |
|---|---|
| System prompt and tool definitions, sent on every call | 3,000 |
| The invoice email and the PDF’s text | 2,500 |
| Model calls per run | 4 |
| Output per call (tool call or summary) | 300 |
| Tool result added after each call | 400 |
Call 1 reads 5,500 tokens. Each later call reads the same 5,500 plus everything added since, so calls 2 to 4 read 6,200, 6,900 and 7,600. That’s 26,200 input tokens and 1,200 output tokens per run. Each extra step adds a longer call than the one before, so a run with eight steps costs more than twice a run with four.
What does one run cost on Claude, OpenAI and Gemini?
Prices per million tokens, checked on 12 November 2026 on Anthropic’s pricing page, OpenAI’s API pricing and Google’s Gemini API pricing:
| Model | Input | Output | Cached read | Per run, no caching | Per run, with caching |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 | $0.0644 | $0.0358 |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 | $0.0322 | $0.0179 |
| Gemini 3.8 Flash, from 1 Jan 2027 | $1.50 | $7.50 | $0.15 | $0.0483 | $0.0260 |
Gemini 3.8 Flash costs $0.75 input and $3.75 output until 31 December 2026, then doubles. We used the 2027 prices because that’s what you’d pay from January.
The small models are much cheaper. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 input and $0.50 output, so the same run costs about $0.003 before caching. Whether a small model handles your invoices accurately enough is something to test on your own documents.
How much does prompt caching save?
Caching can charge a little extra the first time a block of input is stored, then much less each time it’s read again. In this example the 5,500-token start of each run (instructions, tools and the invoice) is written once and read on calls 2 to 4. The growing tail of each conversation is billed at the normal input price.
| Provider | Write price | Read price | Cache lasts | Minimum to cache |
|---|---|---|---|---|
| Claude Sonnet 5.5 | 1.25x input | 0.05x input | 5 minutes, or 1 hour at 2x | 512 tokens |
| GPT-6.1 Terra | 1.25x input | 0.05x input | 30 minutes, or 12 hours at 2.5x | 1,024 tokens |
| Gemini 3.8 Flash | normal input price (implicit caching) | $0.15 per million | not stated | 4,096 tokens |
Sources: Anthropic’s prompt caching docs, OpenAI’s prompt caching guide and Google’s context caching docs. Google turns implicit caching on by default and passes on the saving when a request hits the cache, but doesn’t promise a hit. Explicit Gemini caches also pay a storage price per hour.
With caching, the cost per run fell 44 percent on Claude Sonnet 5.5 and GPT-6.1 Terra and 46 percent on Gemini 3.8 Flash. That counts only reuse within a run.
Reuse across runs depends on timing. The 3,000 tokens of instructions and tools are the same for every run, but a cache entry only helps if the next run arrives before it expires. At 1,000 runs a month spread over working hours, runs are often more than 5 minutes apart, so Claude’s default cache will usually have expired and OpenAI’s 30-minute cache usually won’t. At 10,000 runs a month, both stay warm for most of the day. Check the usage fields in each response (cache_read_input_tokens on Claude, cached_tokens on OpenAI, total_cached_tokens on Gemini) to see what you actually got.
How much human review should you budget?
In this example, review costs more than tokens at both volumes. The formula is runs × share flagged × minutes per review × hourly cost.
Assumptions: the agent flags 20 percent of invoices for a person (new suppliers, changed bank details, amounts over the limit), each review takes 3 minutes, and the reviewer costs $24.36 an hour, the US median wage for bookkeeping, accounting and auditing clerks in May 2025 according to the Bureau of Labor Statistics. That’s wages only, before benefits and overheads.
- 1,000 runs: 200 reviews, 10 hours, $243.60
- 10,000 runs: 2,000 reviews, 100 hours, $2,436.00
Cutting the flag rate from 20 to 10 percent would save more than switching to the cheapest of the three models.
AI agent cost per month at 1,000 and 10,000 runs
| Line | 1,000 runs | 10,000 runs | Source or assumption |
|---|---|---|---|
| API tokens, with caching | $17.90 to $35.80 | $179.00 to $358.00 | Table above, three models |
| Hosting | $24.00 | $24.00 | DigitalOcean Basic Droplet, 2 vCPUs and 4 GiB |
| Monitoring | $29.00 | $29.00 | Langfuse Core, 100k units a month included |
| Human review | $243.60 | $2,436.00 | 20% flagged, 3 minutes each, BLS median wage |
| Upkeep, in-house | $515.52 | $515.52 | 8 hours at $64.44, the BLS median for software developers, QA analysts and testers |
| Upkeep, managed | $1,500.00 | $1,500.00 | Our Basic plan, 1 workflow |
| Total, in-house upkeep | $830 to $848 | $3,184 to $3,363 | |
| Total, managed upkeep | $1,815 to $1,832 | $4,168 to $4,347 |
The model runs at the provider, so one small server covers both volumes here. Langfuse counts each trace, step and score as a billable unit. At about 8 units per run (one trace, four model calls, three tool calls), 10,000 runs is 80,000 units, inside Core’s allowance. The free Hobby plan includes 50,000 units and 30 days of history, which covers 1,000 runs. The 8 hours of in-house upkeep is a guess: 4 hours of changes plus 4 for watching alerts, updating models and rerunning tests.
Which costs do people forget?
- Retries. When a call’s output fails your checks and you run it again, you pay for both attempts, so if 5 percent of calls are rerun, the token line grows by about 5 percent.
- Longer runs. Every extra step re-sends the whole conversation, as the token count section shows.
- Long prompts on Claude Haiku 5.5. A request whose prompt is over 100,000 tokens pays $0.50 input and $2.50 output for the whole request, five times the normal rate.
- Where the data is processed. Keeping inference in the US costs 1.1x on Claude 4.6 and later models, and OpenAI adds 10 percent for regional processing on models released from 5 March 2026.
- Different token counts. Anthropic says its tokenizer for Claude 4.7 and later produces about 30 percent more tokens for the same text than the one before, so token counts from one model don’t carry over to another.
- Price changes and retirements. The Gemini rise on 1 January 2027 is one example. Models also get retired, and the agent has to be moved and retested, which we cover in why AI automations break after launch.
What should you log on every run?
To replace our assumptions with your own numbers, log one row per run with these columns:
run_id, started_at, model, calls, input_tokens, cache_write_tokens,
cache_read_tokens, output_tokens, cost_usd, retries, outcome,
flag_reason, review_seconds, reviewerAfter a month, take the median and the 90th percentile of cost_usd and review_seconds. Budget on the 90th percentile, because a few long runs and slow reviews set the size of the bill.
Should you run it yourselves or have it managed?
If someone on your team can watch alerts, update the model when the provider retires one and rerun the tests after each change, running it in-house is cheaper. If nobody owns that work, it tends not to get done until something breaks.
If you’d rather hand it over, our managed service has three monthly plans: Basic at $1,500 a month for 1 workflow with up to 4 hours of changes, Standard at $3,000 for up to 3 workflows and 10 hours, and Extended at $5,000 for up to 6 workflows and 20 hours. Every plan includes monitoring and alerts, model and prompt updates when your provider changes or retires a model, a rerun of your tests after every change and a written monthly report. There’s a 3-month minimum, then 30 days’ notice to cancel. API tokens, tool seats and hosting are billed to your own accounts at cost, with no markup.
To find out whether an agent would pay for itself in your business, start with the free workflow check.