Agent Money Bench - Adversarial Marketplace: leaderboard
Agent Money Bench (BlockRun, version 1, run 27 October to 11 November 2026) tests whether an agent can be trusted to spend money: it buys the data it needs in USDC on Solana mainnet, answers the question and files an expense report. Unlike Vending-Bench, where the money is simulated, every payment settles on-chain. It has two tracks. Financial Research is 76 open questions answered with live data bought from real providers, graded against per-question rubrics by an AI judge with every score reviewed by hand. Adversarial Marketplace is 80 tasks buying from simulated sellers, some of which raise the price after quoting, let quotes expire, charge twice, deliver the wrong data or hide instructions in what they deliver; it is graded deterministically from the settlement, the answer and the expense report. Each model runs each task once, and the prompt gives no payment rules: it says the money is the client's and that there is a budget. This board is the share of the 80 marketplace tasks completed with a right answer, a matching expense report and no payment error.
Metric: Task success on 80 marketplace tasks (%). Source: agentmoneybench.com. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 98.8 | #6 |
| 2 | Gemini 4 Argon | 98.8 | #11 |
| 3 | Claude Fable 5.2 | 98.8 | #1 |
| 4 | Claude Opus 5.5 | 97.5 | #5 |
| 5 | GPT-6.1 Sol | 97.5 | #10 |
| 6 | Claude Fable 5.1 | 90 | #14 |
| 7 | Beam 1.0 | 76.3 | #189 |
| 8 | GLM-5.3 | 75 | #62 |
| 9 | Claude Sonnet 5.5 | 72.5 | #18 |
| 10 | Kimi K3.1 Pro | 70 | #28 |
| 11 | DeepSeek V4 Pro | 61.3 | #181 |
| 12 | GPT-6 Luna | 45 | #93 |
No result here: #2 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=agent-money-bench-adversarial-marketplace · How It Works · Data refreshed daily, snapshot 2026-11-12.