FinCUABuildBench (Direct Prompting): leaderboard
FinCUABuildBench (Yang et al., 2026) measures whether AI systems can build reliable evaluation tasks for financial computer-use agents: for each of 576 construction requests the system must produce a grounded, executable and replayable episode with a runtime event and a hidden verifier. Qualified task rate in percent, counting only episodes that pass every hard gate, when the model writes the episodes directly without tools; higher is better.
Metric: Qualified task rate (%; share of the 576 preregistered construction slots, 24 financial workflow categories x 8 WorkflowCores x 3 runtime-dynamic profiles, whose generated computer-use episode passes every hard gate: judge acceptance, source and state traceability, execution and replay, verifier pass, valid runtime event, target-capability match and non-duplication, plus every mandatory safety constraint; mean over the preregistered construction runs, macro-averaged over WorkflowCores; the model writes each episode from the construction request without tools). Source: arxiv.org. Saturation forecast: Around 2030. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiMo-V2.5-Pro | 12.3 | #274 |
| 2 | Kimi K3 | 12.2 | #53 |
| 3 | MiniMax-M3 | 11.3 | #265 |
| 4 | GPT-5.5 | 8.7 | #48 |
| 5 | Qwen 3.5 397B A17B | 6.8 | #242 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fincuabuildbench-direct-prompting · How It Works · Data refreshed daily, snapshot 2026-11-12.