← Blog

How Much Does It Cost to Run a Shopping Agent?

June 12, 2026 · Victor Young

A shopping agent's cost has two parts: what you pay to run it, and what you pay when it fails. On Firestarter, the first part is concrete: creating an execution costs 10 tokens and approving it costs 10 more, so a completed purchase runs about 20 tokens. Product search is free. On the Pro plan—$99/month with 10,000 tokens included—that works out to roughly $0.20 per completed purchase at full utilization, with no transaction fees on top.

The second part—the failure bill—is where do-it-yourself agent stacks quietly get expensive, and it deserves most of this article.

The DIY Cost Stack

Suppose you skip the execution layer and build purchasing into your agent directly: a model that reasons about what to buy, plus browser automation to drive merchant checkouts. Your line items look like this:

Inference. Every shopping session burns model tokens on navigation: reading pages, deciding what to click, recovering from unexpected layouts. Checkout flows are long, and agents re-read the page after every step. This is the cost you expected; it is rarely the one that dominates.

Browser infrastructure. Headless browsers per concurrent session, residential proxies (merchant bot-detection blocks datacenter IPs quickly), and CAPTCHA-solving services. None of these are large individually; all of them are recurring, scale with volume, and exist only to make your agent look human to websites that are actively trying to detect it.

Maintenance engineering. The dominant line item. Selector-based automation breaks when merchants ship redesigns, A/B tests, or new bot defenses—which is to say, continuously. Each merchant you support is a small integration you now own, with no contract and no changelog. We have catalogued these failure modes in browser automation vs. commerce API and why AI agents fail at checkout; the summary is that checkout pages are adversarial territory, and you are paying engineers to hold ground that shifts weekly.

Failure cleanup. When a DIY agent fails mid-checkout, the failure is rarely clean: carts abandoned with inventory held, duplicate orders from retries, or—worst—a purchase that succeeded but wasn't what you wanted, now settled with the merchant and yours to unwind through a returns process the agent can't navigate. Each incident consumes human minutes, and human minutes are the most expensive unit in the whole system. (What recovery looks like when it is structured: when AI agent purchases go wrong.)

Reconciliation. Receipts scattered across merchant inboxes, order statuses in per-merchant accounts, card statements that don't map to anything. Someone closes the books monthly; their time is on your agent's bill. The fix is a unified audit trail—which DIY stacks have to build as a second project.

The pattern: DIY converts a per-purchase cost into a standing engineering and operations liability. At low volume it feels free; at real volume the maintenance and cleanup line items dwarf everything.

The Execution-API Model

A commerce execution API collapses that stack into a metered service. The agent sends purchase intents; the execution layer—structured listings, spend limits, approvals, escrowed settlement—does the buying. You stop paying for browser fleets, proxies, and selector repair, because nothing is pretending to be human at a checkout page.

Firestarter's pricing, concretely:

• Free to start: every account gets a one-time 100 tokens plus a 14-day Pro trial—no credit card required. That's enough for about five complete purchases, which is enough to validate the loop end to end.
• Pro: $99/month, 10,000 tokens included. At ~20 tokens per completed purchase, that supports up to ~500 purchases a month. The cap is hard: nothing is ever billed silently. If you run out, top up with token packs (500 for $9.99, 2,000 for $29.99, 5,000 for $59.99) or talk to us about Enterprise.
• Search is free. Agents can browse, compare, and price-check the network without consuming tokens. You pay when you act, not when you look.
• Buyers pay no transaction fees. The token price is the whole buyer-side price; there is no percentage skimmed off your purchase amounts. Sellers list free and pay a 3% commission on completed sales only—which means the network's incentive is completed, non-disputed purchases, same as yours. (Sellers: how to sell to AI agents.)

Two structural notes worth pricing in. Approval steps cost tokens (the 10-token approval charge) but save failure dollars—a declined approval is the cheapest bad purchase you will ever not make. And escrowed settlement means failed or cancelled orders release their authorization rather than entering a refund chase; the cleanup line item that dominates DIY economics mostly disappears because the money never moved. Cancelled executions don't consume additional tokens.

A Worked Example

An ops team runs 60 agent purchases a month—supplies, equipment, client gifts—averaging $85 each.

Execution API: 60 purchases × ~20 tokens = ~1,200 tokens/month, comfortably inside Pro's 10,000. Cost: $99/month flat, about $1.65 per purchase at this volume—and the same $99 covers them up to ~500 purchases as the agent takes on more. No fees on the ~$5,100/month of purchase volume. Approval queue on a phone; receipts and tracking in one audit trail; books close from one export.

DIY: inference ($20–60/month at this volume), proxies and CAPTCHA solving ($50–150/month for reliable residential coverage), and the real costs: an engineer's recurring hours when merchants change layouts, plus cleanup on the orders that fail or mis-fire. At a 5–10% incident rate—optimistic for selector-driven checkout—that's 3–6 incidents a month, each eating thirty-plus human minutes and occasionally a disputed charge. The infrastructure is cheap; the ownership is not.

The crossover logic is simple: DIY's marginal cost per purchase can look lower, but its fixed cost is an engineering commitment that never ends. Unless driving arbitrary storefronts is itself your product, you are buying the wrong problem.

How to Estimate Your Own Bill

1. Count your monthly agent purchases (be realistic—count what a human does today that you want delegated).
2. Multiply by ~20 tokens; compare against the one-time 100 free tokens for validation, then Pro's 10,000/month.
3. Price your failure tolerance: every purchase that needs human cleanup costs more than the purchase's token price by an order of magnitude. Weight any alternative by its incident rate, not its sticker price.
4. Run the free trial against a week of real purchases and measure, rather than modeling.

FAQ

What does a single purchase actually cost in tokens?

Creating an execution costs 10 tokens; approving it costs 10 more—about 20 tokens for a full purchase. Product search is free, so comparison shopping costs nothing until the agent acts.

Is there a percentage fee on my purchase amounts?

No. Buyers pay no transaction fees—tokens are the entire buyer-side cost regardless of order value. Sellers pay a 3% commission on completed sales only.

What happens when I hit the monthly token cap?

Executions pause—nothing is ever billed silently. Top up with a token pack (500 for $9.99, 2,000 for $29.99, 5,000 for $59.99), upgrade, or wait for renewal.

Do failed or cancelled purchases consume tokens?

Cancelled executions don't consume additional tokens, and because settlement waits for delivery confirmation, a cancelled order releases its payment authorization rather than starting a refund process.

Can I try it without paying?

Yes. Every new account starts with a one-time 100 free tokens plus a 14-day Pro trial—no credit card required. That's roughly five complete purchases, enough to test search, approval, and delivery end to end. Pricing details here.