Backed by HF0

Agents spend money. We measure, and make them better at it.

9 of 12 frontier models paid a seller who was cheating them.

Payments graded from the settlement record. The method and example payment traces are public.

LeaderboardOct–Nov 2026
Agent Money Benchmark · adversarial track · selected rows
#ModelTask successScam rate
1Gemini 4 Argon98.8%0.8%
6Claude Fable 5.190.0%2.4%
9Claude Sonnet 5.572.5%8.0%
10Kimi K3.1 Pro70.0%18.3%
11DeepSeek V4 Pro61.3%31.4%

Capable is not the same as safe: Claude Fable 5.1 answers as many research questions as Claude Opus 5.5, yet made money errors on 5% of tasks.

Agents buy data with a client’s money; some sellers cheat. Payments graded from the ledger. Method ↗

The loop

One loop, from a failed payment to a better agent.

Agents pay for models, data and tools through our endpoint, so we see the payments they get wrong. Each kind of mistake becomes a test we synthesize. Tests become training environments. What we learn goes into the environments labs train on, and next into a model of our own for payment decisions.

  1. See

    We run the endpoint agents pay through, so we know where payments break in practice. Every task we test or train on is synthesized by us; no customer data goes in.

    01
  2. Measure

    Public and private tests graded from the ledger: right party, right amount, no scams, within limits, honest report.

    02
  3. Train

    Environments built from the same tests, for labs training agents that handle money.

    03
  4. Run

    The endpoint agents already call. ClawRouter, open source and on your side, can pick the model for each request.

    04

Three ways in

Financial institutions

Before an agent touches your money.

We run your own payment workflows (invoice approval, vendor bank-detail changes, duplicate bills, spend under a budget) across models and setups, and check every payment against the ledger. You learn which setup you can trust, what it costs per completed task, and where it fails.

Verified Eval →

AI labs

Environments for agents that handle money.

The reward comes from what settled, not from a judge. Every task is synthesized by us. The benchmark keeps a held-out set we never publish.

For AI labs →

Developers

Every model and tool, one endpoint.

125 models, each call priced before it runs. Pay from a wallet with no account, or get an API key and pay by card or invoice.

For developers →

Build with BlockRun

One infrastructure layer for AI teams and developers.

For enterprise

Prove your agents handle money right, and what that costs.

Talk to our team

For developers

Start building.

Explore the docs