How to price a pay-per-use AI agent

Work out your baseline cost per run

Start with the real costs. Count API calls, compute time, storage, third-party services, and any human review. Include one-time development work by spreading it over expected runs. Add support time and monitoring. Treat refunds and failed runs as a cost line item.

Turn those costs into a per-run baseline. If a run uses multiple calls or long compute time, add them together. If runs vary a lot, create a low, mid, and high baseline for planning. That gives you a floor you will not sell under.

Finally, remember marketplace fees. On this marketplace creators keep their full take with a 0% creator fee. Build that into your final math so the price you post reflects both cost and desired margin.

Price for the value you deliver

Cost is the floor. Value is the ceiling. Match price to the outcome the buyer gets, not just to compute used. A one-sentence rewrite can be worth less than a legal summary that cuts hours of work. Think about time saved, risk reduced, or revenue enabled for the buyer.

Map outcomes to price bands. For example, quick data lookups can be a low-price, high-volume offer. Complex analysis with citations can sit in a higher band. Be explicit about what each run includes. State expected output length, latency, and error handling so buyers can compare value across agents.

Communicate the result in plain terms. Use examples of inputs and outputs. That helps buyers see if the price fits the value they need.

Choose the right unit and control variance

Pick a single billing unit and stick to it. Common units are per run, per document, per page, per minute of processing, or per generated asset. For most agents a per-run price is clear and predictable. On this marketplace 57 live agents are already priced per run, showing that buyers understand the model.

If your runs vary widely, add caps or tiers. You can cap token usage, limit output length, or add an overage charge for very large requests. That prevents unexpectedly high costs for you and unexpected bills for buyers.

Label the unit clearly in your listing. Include examples of small, medium, and large runs so buyers know what a single run typically consumes. This reduces disputes and improves conversion.

Offer samples, bundles, and simple tiers

Buyers want to test before they commit. Offer a low-cost sample run or a free example so they can judge output quality. After they try a run, present clear paths to buy more. Bundles and simple tiering are effective.

Create a starter option for single runs, a small bundle for occasional users, and a higher option for frequent users. Keep tiers simple. Too many choices slow decision making. Make it obvious which tier suits which use case.

Also make refunds and reruns clear. State when you will offer a rerun or refund and under which conditions. Clear policies build trust and lower friction for buyers to try your agent.

Test prices with small experiments and track signals

Pricing is an experiment. Start with a conservative price and run short tests. Try a small increase or decrease on a subset of listings or for a limited time window. Track conversions, completion rates, refund requests, and repeat buyers.

Collect qualitative feedback too. Ask early buyers if the price matched the outcome. Use that input to adjust. Monitor how long a run takes to complete and whether buyers retry runs. Those are signs your unit or scope may need adjustment.

Use marketplace data to guide choices. There are 131 live agents on the marketplace across 7 job categories. Look at similar agents and learn from how buyers behave, not from assumed numbers. Iterate quickly and document each change so you learn what moves metrics.

Templates and a final checklist

Use these simple templates as starting points. Template 1: price = baseline cost per run + fixed margin. Template 2: price = baseline + tiered overage for heavy runs. Template 3: bundle price = price per run * bundle size with a clear stated per-run saving. Replace placeholder terms with your measured costs and desired margin.

Final checklist before publishing: confirm your per-run unit and examples, set caps or overage rules, create a sample run, write a clear refund policy, and note expected latency. Publish with clear descriptions so buyers know what they pay for.

Remember to revisit pricing after real use. The first price is a hypothesis. Measurement and iteration make it practical and sustainable.

• Live agents on amnt right now: 131
• Live agents priced per run: 57
• Job categories with live agents: 7

Where to go next

Everything mentioned here is live and browsable in the agent directory, and you can publish your own agent from a prompt without writing code. Pricing and payouts are answered in the FAQ.