AgencyBench V2 - Research: leaderboard
Evaluates autonomous agent task completion using AgencyBench V2 rubric scoring across Game, Frontend, Backend, Code, Research and MCP tasks. This series reports Research average task score on the published V2 agent setup. It is distinct from V1, average attempts and comparison deltas. Source: the authors’ public leaderboard data.
Metric: Average Task Score (%). Source: agencybench.opensii.ai. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 71.4 | #148 |
| 2 | Claude Opus 4.5 | 68.8 | #87 |
| 3 | GPT-5.2 | 64.4 | #115 |
| 4 | Grok 4.1 Fast | 63.8 | #227 |
| 5 | Gemini 3 Pro | 40 | #82 |
| 6 | Qwen 3 235B A22B (Thinking) | 34.4 | #323 (Qwen 3 235B A22B) |
| 7 | Kimi K2 (Thinking) | 33.6 | #252 (Kimi K2) |
| 8 | GLM-4.6 | 32 | #265 |
| 9 | DeepSeek V3.2 | 22 | #210 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agencybench-v2-research · How It Works · Data refreshed daily, snapshot 2026-11-05.