AgentMeter - Task Quality: leaderboard
Command-line agent tasks from Terminal-Bench, SkillsBench and DA-Code (the 30-task AgentMeter Core Subset, Chi et al., 2026), each run once by a language model inside one agent harness (Claude Code, Codex CLI, Kimi Code CLI or Qwen Code) with the harness's default settings. Task quality is the mean evaluator reward from 0 to 1, with partial credit kept; higher is better. Each entry is a model and harness pair, because the paper finds that the best harness differs from model to model.
Metric: Task quality: mean evaluator reward (0-1; 30-task Core Subset of Terminal-Bench, SkillsBench and DA-Code tasks, 10 per effort stratum, one run per LM-CLI configuration, partial credit kept). Source: arxiv.org. Saturation forecast: Around February 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.63 | #158 |
| 2 | GLM-5.1 | 0.62 | #175 |
| 3 | Gemini 3.5 Flash | 0.62 | #98 |
| 4 | GPT-5.3 Codex | 0.6 | #114 |
| 5 | Qwen 3.6 Plus | 0.57 | #217 |
| 6 | DeepSeek V4 Pro | 0.56 | #181 |
| 7 | MiniMax-M2.5 | 0.49 | #590 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentmeter-task-quality · How It Works · Data refreshed daily, snapshot 2026-11-12.