CalBench - Varied Costs - Excess Calibrated VPS: leaderboard
Unnecessary inferred private-calendar disclosure beyond completed coordination, in slot-equivalent units; lower is better. Canonical mixed-agent calendar negotiation, paper v3 Table 1. Uniform and varied displacement costs are separate settings; 45 or 135 evaluated agent-game seats per model, not homogeneous teams. Means retain original units. This series reports Excess calibrated VPS (slot-equivalent units). Source: https://arxiv.org/html/2605.09823v3.
Metric: Excess calibrated VPS (slot-equivalent units). Source: arxiv.org. Saturation forecast: Not forecast. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick Instruct | 3.49 | #467 |
| 2 | GPT-5.4 Mini | 4.6 | #228 |
| 3 | Gemini 3 Flash (Preview) | 7.21 | #90 |
| 4 | Gemini 3.1 Pro (Preview) | 8.1 | #64 |
| 5 | Qwen 3.6 Plus | 8.56 | #124 |
| 6 | Claude Sonnet 4.6 | 12.85 | #94 |
| 7 | DeepSeek V4 Pro | 15.27 | #112 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=calbench-varied-costs-excess-calibrated-vps · How It Works · Data refreshed daily, snapshot 2026-11-05.