When Simulation Lies: leaderboard
Tool-use language agents are evaluated on benchmarks that assume clean inputs, unambiguous tool registries, and reliable APIs. Real deployments violate all these assumptions: user typos propagate into hallucinated tool names, a misconfigured request timeout can stall an agent indefinitely, and duplicate tool names across servers can freeze an SDK. We study these failures as a sim-to-real gap in th
Metric: Pert Acc (imported). Source: benchmarklist.com. Saturation forecast: Around April 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 14B (Non-reasoning) | 52.92 | #537 (Qwen 3 14B) |
| 2 | Qwen 3 32B (Non-reasoning) | 52.36 | #443 (Qwen 3 32B) |
| 3 | O4 Mini | 50.2 | #180 |
| 4 | Qwen 3 8B (Non-reasoning) | 44.83 | #674 (Qwen 3 8B) |
| 5 | DeepSeek R1 Distill Qwen 14B | 44.24 | #844 |
| 6 | Qwen 3.5 9B | 43.7 | #379 |
| 7 | Qwen 2.5 3B Instruct | 39.01 | #1162 |
| 8 | Qwen 2.5 1.5B Instruct | 36.68 | #1407 |
| 9 | Llama 3.2 3B Instruct | 34.1 | #1352 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=when-simulation-lies · How It Works · Data refreshed daily, snapshot 2026-10-20.