What an AI worker actually costs to run

AI worker pricing runs on three meters: the plan, the tool calls and the model tokens. What each measures on WorkerKit, with real figures and worked examples.

"What does an AI worker cost?" has three answers, because three meters are running. Most confusion about agent pricing comes from mixing them up.

The third is the one people underestimate, and the one most platforms mark up quietly. Take them in order.

The plan buys capacity

Four things in that table are worth pulling out.

Free is not a trial. No card, no expiry clock. Five workers and 500 tool calls a day runs real jobs, and the complete safety layer is included rather than held back.

Team is flat, not per seat. $299 covers up to 10 users, so the tenth teammate costs nothing extra. The tool-call allowance is pooled across the account rather than divided, so one heavy worker does not starve the rest by design.

Yearly is not a discount. Yearly is twelve times monthly. The incentive is a wallet load on day one: $50 of runtime credit on Pro, $500 on Team. That credit spends on hosted runs and BYOK platform fees alike, is drawn before money you load yourself, and stays good for 12 months.

The safety layer is never tier-gated. The app firewall, redaction, contact rules and one key with its own scope list per worker ship on every plan, Free included.

Tool calls are the capacity you actually spend

A tool call is one action in a connected app: reading a thread, sending a reply, adding a calendar event, looking up a CRM record.

One run makes several. A worker triaging an inbox might list threads, read four, check two calendar slots and write one draft. Eight tool calls in a single run.

So the arithmetic that matters is calls per run multiplied by runs per day:

Frequency is the multiplier on everything, which is the real argument for choosing a schedule that matches the job rather than the fastest one available. A minute-by-minute schedule is usually a webhook trigger nobody has wired yet: it burns calls all day to find nothing, and still reacts up to a minute late.

The caps are hard. At the cap a worker stops until the counter resets, then picks its schedule back up. Nothing is deleted, no setting changes, and no overage lands on a bill you did not agree to. For unattended work a hard cap is the better failure mode: a runaway loop on a soft cap is an invoice, on a hard cap it is a paused worker and a receipt explaining why.

Model tokens are a pass-through

Tokens bill at the provider list price, to the cent, with no markup. Two ways to pay.

The wallet. A prepaid balance that hosted runs draw on. Loading it costs 5.5% with a $1 minimum, charged on top, and the credited balance is exactly what you loaded:

The $1 minimum dominates small loads and the percentage takes over as they grow. Nothing is skimmed off the balance itself.

Your own key. Add an Anthropic, OpenAI, Google or xAI key and those tokens bill straight to your provider account. The first **1M tokens each month carry no platform fee**; past that it is 5% of list, settled from the wallet. Removing a key returns that provider to wallet billing.

Under 1M tokens a month the two cost close to the same thing. What BYOK buys at low volume is control rather than savings: existing committed spend, one bill for finance, a direct relationship on rate limits and quota.

Why no markup is a structural claim

It would be easy to take 20% of tokens and call it platform margin. Not doing so is what makes two other promises real.

WorkerKit is model agnostic: Claude, ChatGPT, Gemini and Grok all drive the same kits, and you can send a single run to a different model to compare them on the same job. If tokens carried a markup, the platform would earn more when you picked the bigger model, and "switch any time" would sit in tension with its own revenue.

It also means nothing in the pricing rewards a chatty worker. A kit that does the job in fewer tokens is straightforwardly better, and costs WorkerKit nothing to encourage.

WorkerKit's fees are the plan price and the wallet top-up fee. Both are stated in the open, and that is the whole list.

Reserve before run

One behaviour to know before leaving a fleet running: a run reserves its token budget before it starts.

A balance that will not cover the reserve skips the run and says so on the receipt. Top up, and the next due run goes ahead. Nothing is half-completed and no setting changes.

The alternative, discovering mid-run that the balance ran out, leaves you unpicking work that got halfway through a mailbox. For an unattended system, predictable refusal beats partial success.

Three real workers, costed

Using real kits from the directory, with their actual apps:

All three fit inside Free's 500 a day, together. That is the point of the free plan being 5 workers rather than one: a small real fleet fits.

What pushes you to Pro is usually not one heavy worker but **frequency and headcount**: a support triager every 15 minutes, plus a lead qualifier on an event trigger, plus six more workers, is a Pro-shaped account.

Model tokens for those three are separate and depend on the model and how much context each run reads, billed at list price from the wallet.

Where this pricing model does not suit you

Every other decision on this site gets a "when not to", and pricing should not be the exception just because it is ours.

• You want one predictable invoice. Two meters means the plan is fixed and the tokens are not, so a heavy month costs more than a light one. If finance needs a flat number, the wallet's variability is friction, and BYOK moves it rather than removing it.
• Your usage is spiky and small. The wallet's $1 minimum load fee is 20% on a $5 top-up. Occasional light use pays disproportionately unless you load in larger amounts less often.
• You need one worker, run once a week. Below a certain volume the honest advice is that a person does the job faster than you will configure it, and no plan price fixes that.
• You need self-hosting today. WorkerKit hosts every run; your own machine or infrastructure is coming later. If that is a hard requirement now, the pricing question does not arise.
• You wanted tokens bundled. Some platforms fold model cost into a seat price, which is genuinely simpler to budget. The objection below takes that case seriously rather than dismissing it.

The reasonable objection

"Separating token cost from the plan transfers pricing risk to me. If model prices rise, my bill rises and yours does not. A platform that bundles tokens into a seat price is absorbing that risk on my behalf, and charging for it. You have unbundled the risk and called it transparency."

That is the sharpest version and the first half is simply true. Pass-through pricing means provider price changes reach you directly, in both directions, and a bundled competitor is running that hedge for you.

Three things are worth weighing against it, and the third is the real answer.

Provider prices have generally fallen rather than risen as models improve, so the risk being transferred has historically been upside. That is an observation about the past and not a promise.

A bundle prices the hedge, and it prices it for the average customer. If your workers are light, you are paying for somebody else's heavy usage. Whether the hedge is worth it depends entirely on which side of the average you sit, and most people cannot tell without the itemised number that bundling removes.

And the incentive point cuts deeper than the price. A platform earning margin on tokens earns more when your worker reads more context and picks a larger model. That interest shows up not in the price list but in defaults, in which model is recommended, and in how hard anyone works to make a kit efficient. No-markup is worth less to you as a discount than as the removal of that interest.

The fair conclusion: if you want one predictable number and are willing to pay for the hedge, a bundled competitor is a legitimate choice and this model will annoy you. The trade is real rather than one-sided.

What the honest answer looks like

For a single daily worker: **the Free plan, plus your model's list price for the thinking.** Nothing in between, and no per-seat charge until you are a team.

For a working fleet: $29 a month, plus tokens at list price, plus a wallet top-up when you choose to load it.

Full figures are on /pricing, and the same numbers in markdown for agents are at /pricing.md.

FAQ

Is the WorkerKit free plan a trial?

No. Free is $0 with no card and no expiry: 5 workers, 500 tool calls a day, one schedule per worker, and the complete safety layer. It is a real plan you can run a small fleet on indefinitely.

Does WorkerKit mark up model tokens?

No. Model tokens bill at the provider's list price to the cent, whether they are paid from the prepaid wallet or through your own provider key. WorkerKit's own fees are the plan price and the wallet top-up fee of 5.5% with a $1 minimum, both stated openly.

What happens if I hit the daily tool call limit?

The worker stops until the counter resets, then picks its schedule back up. Nothing is deleted and no setting changes. The caps are hard by design, so an unattended fleet cannot generate an overage bill you did not agree to.

Is BYOK cheaper than using the wallet?

Not necessarily. Under 1M tokens a month both routes bill at list price with no platform fee, so the difference is control rather than cost: your own committed spend, a single provider bill, and a direct relationship on rate limits. Above 1M tokens the BYOK platform fee is 5% of list.

What do I get for paying yearly?

A wallet load on day one: $50 of runtime credit on Pro, $500 on Team. Yearly is twelve times the monthly price, so the incentive is the credit rather than a lower rate. It spends on hosted runs and BYOK platform fees, is drawn before money you load yourself, and lasts 12 months.

How many workers can I run on each plan?

Five on Free, 50 on Pro, and 500 on Team. Team is flat for up to 10 users rather than per seat, and its 50,000 daily tool calls are pooled across the whole account.