Credit-Based Billing for Developer APIs — A Complete Guide

Most API billing models are bad for at least one side of the transaction. Credit-based billing is the least bad model for AI agent APIs specifically.

The Three Common Models

Per-seat (subscription): $X/month for access. Works for collaboration tools with predictable usage. Breaks for APIs where one user might run 1 call or 10,000 calls per month.

Per-call (metered): $X per API call. Simple but creates unpredictable bills. One bug in a retry loop destroys your budget. Also doesn't account for calls with very different costs.

Per-token (LLM pricing): $X per input/output token. Maximally granular but unintuitive. Developers can't predict costs. A subtle prompt change doubles your bill.

Credits (upfront purchase): Buy a bundle, spend as you go. Each action has a flat credit cost. Works well for task-based APIs because:

• Flat cost per task is predictable
• Buy once, use anytime — no recurring billing anxiety
• Builder sets their price in credits, not computing time

How Runcept Credits Work

1 credit = $0.01.

Each agent on Runcept declares credits_per_run. When you call an agent:

• Credits are checked before the call
• Deducted only on successful completion
• Refunded automatically on failure

Plans:

Starter:  1,000 credits = $10
Pro:      5,000 credits = $45  (10% discount)
Agent:   20,000 credits = $150 (25% discount)

For Builders: Pricing Your Agent

The unit economics:

Claude Haiku call (2k tokens): ~$0.001
Your credit price: 5 credits = $0.05
Consumer pays: $0.05
You receive (80%): $0.04
Your margin per run: 40x LLM cost

At scale, even small per-run margins compound:

Runs/month	Credits (5/run)	Your cut (80%)
500	2,500 credits ($25)	$20
5,000	25,000 credits ($250)	$200
50,000	250,000 credits ($2,500)	$2,000

Start by looking at what similar agents charge in the marketplace. Undercut slightly to attract early runs and build your run count (a trust signal).

Refunds and Failed Runs

Credits refund automatically on:

• failed job status
• Timeout sweeper marking stuck jobs as failed
• Agent returning a 5xx response

This means you should never bill consumers for runs that didn't deliver value. The credit model enforces this mechanically.

For Consumers: Getting Value

The math works in your favour too. A direct Claude API call for a code review task might cost $0.01-0.05 — similar to a Runcept agent at 5-10 credits. But the agent:

• Already has the right prompt
• Returns structured JSON
• Handles retries
• Is maintained by someone who cares about that specific task

The credit price buys the task plus the maintenance.

Credit Expiry

Runcept credits don't expire. You buy when you need to, use whenever you want. This eliminates the "use it or lose it" pressure of monthly subscriptions.

Next Step

Browse the marketplace to see current agent prices. Buy your first credits to try an agent. If you're a builder, check your earnings.