The price map

The current agentutility registry has 254 priced endpoints across eight clusters: edge-finance, locale, edge-market, prooflayer, wordmint, web-probe, mediakit, and synthforge.

The cheapest call is $0.001. The highest is $0.30.

The median is $0.01.

The most common sticker price is $0.005. That matters for routers. If an agent sees agentutility as a paid tool catalog, it shouldn't treat the catalog as a set of equal-cost choices. The price curve has shape.

Here’s the bucket view:

| Price bucket | Endpoint count | | --- | ---: | | $0.001 to under $0.005 | 46 | | $0.005 to under $0.01 | 48 | | $0.01 to under $0.02 | 40 | | $0.02 to under $0.05 | 77 | | $0.05 to under $0.10 | 16 | | $0.10 to under $0.20 | 17 | | $0.20 to under $0.30 | 9 | | $0.30 | 1 |

So the center of gravity is obvious. Most of the registry sits from $0.005 through $0.04.

Most calls are cheap enough to make during planning, validation, enrichment, or retries. The pricey end exists, but it’s thin.

Why $0.005 to $0.05 is crowded

Agent tools have weird unit economics. A lot of the work is small but high-value: parse a VIN, check a domain, normalize a timestamp, convert a file, fetch an SEC item, score a package risk, turn HTML into Markdown.

These aren’t subscription products. They’re one-shot answers.

For an LLM router, $0.005 is the “don’t ask the human” tier. If the call can prevent a bad answer, add a citation, verify a claim, or fill a missing field, it’s cheap enough to run. Think hash-string, slugify, jwt-decode, cron-next, country-info, or dns-lookup.

At $0.01, the call can still be routine. The median lands there because most endpoints return a single typed answer an agent can act on without ceremony. The agent pays a cent, gets that answer, and moves on.

The $0.02 to $0.04 band is the busiest bucket in the registry. That’s where calls start doing more work. Web fetches. Document parsing. Market lookups. Security checks. Geo queries. The price is still small, but it says the endpoint may touch an upstream API, spend more CPU, or run a model.

$0.05 is a line. Past it, the agent should ask a better question before paying.

The tail is real

There’s one $0.30 endpoint: image-generate-ultra.

That’s the far end of the chart, and it’s there for a reason. Image generation can burn paid model capacity in a way that cron-next never will. A router shouldn’t call it as a casual side effect. It should call it when the user asked for an image, when the prompt is ready, and when the output has a place to go.

The $0.20 tier has eight endpoints. The $0.25 tier has one.

That upper band is small enough to handle with explicit routing rules:

  • Don’t call above $0.10 during background browsing.
  • Don’t retry above $0.10 without checking the failure reason.
  • Don’t batch high-price media calls unless the user asked for multiple outputs.
  • Do cache results when input hashes match.

Cheap calls can be speculative. Expensive calls need intent.

What agents should do with this

A router can treat price as a first-class planning signal.

A simple policy works:

| Task state | Suggested cap | | --- | ---: | | Exploring possible tools | $0.005 | | Filling missing structured fields | $0.01 | | Verifying a claim or artifact | $0.02 | | Calling web, market, or document endpoints | $0.05 | | Generating media or running heavy LLM work | User intent required |

That policy still leaves most of the catalog open. It blocks accidental spend where mistakes hurt.

And it matches the registry. The catalog isn’t priced like SaaS seats. It’s priced like agent actions on Base mainnet, where each HTTP 402 response can quote a job, receive payment in USDC, and return a machine-readable result.

The point for builders is simple: fixed per-call prices make routing possible. The point for agents is sharper: price tells you when to think twice.