Guide · Pricing
How much do AI agent tool calls cost? Per-call, per-task and per-seat pricing compared
An AI agent's tool calls cost whatever the unit of your integration layer counts, multiplied by how often your agent triggers it. The common units are the seat, the task (one workflow step that ran), the tool call (one request from the model) and the upstream call (one request actually sent to a vendor). actionlake, an MCP gateway, bills the upstream call: under a tenth of a cent each on every paid plan, 100,000 a month free, with cache hits, searches, dry runs and failures you did not cause (a vendor's 401, 403, 429, 5xx or a timeout) free. The rest of this guide shows how to compare units, with worked costs.
Four units, four different counts
The same agent session produces very different numbers depending on what is counted. Take one question: "Which invoices are overdue?" The agent searches for an action, reads its schema, dry-runs the call, then fetches three pages of invoices, one of which the cache already holds.
| Unit | What it counts | Count for the example |
|---|---|---|
| Seat | People with access, whatever they do | 1 seat, whether they ask once or a thousand times |
| Tool call | Every request the model makes to the integration layer | 4 (search, schema, check, fetch) |
| Task or operation | Each workflow step that ran | Depends on how the step is defined; often each action that succeeded |
| Upstream call | Requests actually sent to the vendor that returned an answer | 2 (the cached page is free) |
None of these is wrong. They price different things. What matters is how the unit grows with your use, and what it charges for that is not work done in the vendor's system.
Per seat
Seat pricing charges per person with access. It is predictable for a fixed team and stops making sense when agents, not people, do the work. One person's agent can make thousands of calls a day; a scheduled plan has no seat at all. If an integration layer prices per seat, ask how it prices an unattended agent, a schedule and a shared service account.
Per task or operation
Workflow automation platforms usually bill a task or an operation: each step of a flow that ran. The unit was designed for flows a person built in advance. Agents differ in two ways: they explore before they act (searches, schema reads, dry runs), and they read far more than they write. Check whether exploration counts as tasks, whether a read of many pages is one task or many, and whether failed steps count.
Per tool call
Some MCP gateways and hosted servers bill each tool call the model makes. This counts the model's chatter as well as the work: a model that searches twice before acting pays for both searches. It also counts calls that never reached a vendor, such as a refusal or a validation error, unless the provider says otherwise.
Per upstream call
actionlake bills the upstream call: one HTTP request sent to a vendor that returned a definitive, attributable response. A tool call is not a billable unit. Only two of the eight tools, do and ask, can send upstream calls at all.
| Free in actionlake | Why |
|---|---|
help, find_actions, get_action, check, cost, account | Nothing is sent to a vendor |
| Cache hits | Answered without calling the vendor |
| Calls the vendor answered with 401, 403, 429, a 5xx or a timeout | No definitive answer was delivered |
| Calls refused by validation, permissions or a cap | Nothing left actionlake |
The join in ask | Our compute, not a vendor's |
| Idempotent replays | The first answer is returned again |
Cache hits have a fair-use line: 200 hits per billed call, then US$0.02 per 1,000 hits. Every response carries its own meter in the AL-Calls, AL-Micros and AL-Cache headers, so you can see the cost of each call as it happens.
actionlake's plans
From the price book, version 2026-10-24. Each region is priced and invoiced in its own currency; prices exclude tax.
| Plan | US$ a month | A$ | NZ$ | Calls included | Then, per 1,000 |
|---|---|---|---|---|---|
| Free | 0 | 0 | 0 | 100,000 | Hard stop |
| Starter | 49 | 79 | 89 | 250,000 | 0.30 |
| Pro | 149 | 229 | 259 | 750,000 | 0.28 |
| Team | 399 | 599 | 699 | 2,000,000 | 0.25 |
| Business | 1,199 | 1,799 | 2,099 | 5,000,000 | 0.20 |
| Enterprise | from 3,000 | from 4,500 | from 5,400 | 15,000,000 | 0.18 |
Business and Enterprise are annual agreements, invoiced monthly. The per-1,000 numbers are the same numerals in each currency. On every paid plan a call costs under a tenth of a cent; on Enterprise overage it is 0.018¢. People and agents are never counted: there are no seats on any plan. Plans differ in connections, enabled schedules, how long run traces are kept, the Dedicated IP, the availability target and support.
Worked costs, in US dollars
| Monthly upstream calls | Cheapest plan | Monthly cost | Effective cost per 1,000 |
|---|---|---|---|
| 3,000 | Free | $0 | $0 |
| 100,000 | Free, at its hard stop | $0 | $0 |
| 400,000 | Starter: $49 + 150,000 × $0.30 / 1,000 | $94 | $0.235 |
| 1,000,000 | Pro: $149 + 250,000 × $0.28 / 1,000 (Starter would be $274) | $219 | $0.219 |
| 3,000,000 | Team: $399 + 1,000,000 × $0.25 / 1,000 | $649 | $0.216 |
| 10,000,000 | Business: $1,199 + 5,000,000 × $0.20 / 1,000 | $2,199 | $0.220 |
Starter and Pro cost the same at 583,333 calls a month, Pro and Team at 1,642,857, and Team and Business at 6,000,000. At 10,000,000 a month the Business 10M bucket prices the same month at $2,100. Above 15,000,000 a month, talk to us.
The hidden multipliers
Whatever the unit, four things multiply what an agent costs:
- Pagination. A list of 2,000 invoices at 100 a page is 20 requests. Ask whether each page is billed. In actionlake each page is one upstream call, and
asklets you bound a step withmax_pages. - Retries. A loop that retries a failing call can bill each attempt. actionlake does not bill a vendor's 401, 403, 429, 5xx or a timeout, and never retries a write. A request the vendor rejects as invalid, such as a 400 or 404, is billed.
- Exploration. Agents search and read schemas before they act. If those are billed, a careful agent costs more than a careless one.
- Model tokens. The integration bill is often smaller than the model bill. An agent that fetches 2,000 rows into its context to answer a yes-or-no question pays for every token.
askjoins rows in the gateway and returns only the answer, so the rows never reach the model. See cross-system questions.
How to cap what an agent can spend
Pricing is half the story; the other half is a ceiling. In actionlake each key can carry cap_calls, a hard limit in upstream calls, refused before the call is sent. Caps cascade: a child key's calls count against every key above it. The Free plan stops at 100,000 with no overage. The cost tool shows spend so far, the forecast and each key's headroom, free. See agent spending cap.
Questions to ask any provider
| Question | Why |
|---|---|
| What exactly is one billable unit? | Tool calls, tasks and upstream calls differ several-fold |
| Are searches, schema reads and dry runs billed? | Agents make many |
| Are failed vendor calls billed? | Outages should not cost you |
| Are cached answers billed? | Agents repeat reads |
| Is each page of a list billed? | Lists multiply |
| Is a cap enforced before or after the call? | After means the crossing call still costs |
| How much notice before a price rise? | actionlake's price book commits to 60 days |
Frequently asked questions
What does one AI agent action cost on actionlake?
Within a plan's allowance, the monthly fee covers the calls; paid plans start at Starter, US$49 a month for 250,000. Past a paid plan's allowance, an upstream call costs from US$0.00018 to US$0.0003, depending on the plan. Free includes 100,000 calls a month and stops there. Searches, dry runs, cache hits and failures you did not cause (a vendor's 401, 403, 429, 5xx or a timeout) cost nothing.
Is per-call pricing cheaper than per-task pricing?
It depends on what each unit counts and on your mix of reads and writes. Count a week of real agent activity in both units, including exploration and pagination, and compare. A task and an upstream call are not the same unit, so compare totals, not unit prices.
Do I pay for the AI model separately?
Yes. actionlake does not run or resell a model. You pay your model provider for tokens. ask reduces those tokens by returning answers instead of rows.
Can I prepay?
Yes. Prepaid blocks are billed at the Starter rate, carry a bonus on larger blocks, and never expire. See pricing.
Try it on your own systems.
100,000 free calls a month. No card. Hard caps on every key.
