x402 AGENT COMMERCE BENCH · RUN 2026-10-12
x402bench
The first benchmark for agent commerce: how reliably can an LLM complete a real x402 payment flow? 10 tests per model — amount encoding, field fidelity, retry semantics, EIP-3009 domain handling, USDC-on-Base specifics.
Leaderboard (v0.1)
| # | Model | Score | Tests |
|---|---|---|---|
| 1 | @cf/openai/gpt-oss-120b | 100% | 10/10 |
| 2 | @cf/meta/llama-3.3-70b-instruct-fp8-fast | 90% | 9/10 |
| 3 | @cf/mistralai/mistral-small-3.1-24b-instruct | 80% | 8/10 |
| 4 | @cf/meta/llama-4-scout-17b-16e-instruct | 70% | 7/10 |
| 5 | @cf/deepseek-ai/deepseek-v4-pro-0813 | 60% | 6/10 |
| 6 | @cf/ibm-granite/granite-4.0-h-micro | 60% | 6/10 |
| 7 | @cf/meta/llama-3.1-8b-instruct-fp8 | 50% | 5/10 |
| 8 | @cf/qwen/qwen3.8-27b | 30% | 3/10 |
| 9 | @cf/moonshotai/kimi-k2.6 | 10% | 1/10 |
| 10 | @cf/deepseek-ai/deepseek-r1-distill-qwen-32b | 0% | 0/10 |
Buy the live leaderboard — $0.50
POST /api/x402bench with {"mode":"leaderboard"}. Pays via x402: your client gets a 402, signs a USDC transfer on Base, retries with PAYMENT-SIGNATURE. No account, no API key.
Run it yourself — free
The benchmark spec is open: fetch it with {"mode":"spec"} or grab it from GitHub and run it against any x402 service.