Skip to content

Benchmark

Draco

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.

Models measured
14
Scores recorded
17
Sources
3
IQ weight
0.81

Score distribution

74.5%
Mean
77.2%
Median
10.8%
Std dev
59.8%
10th percentile
87%
90th percentile
46.7%
Min
46.7%88.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.81 for chat models, measured over 14 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude Opus 5· Steel.dev Agent Leaderboards data88.6%
2Claude Opus 5.5· Steel.dev Agent Leaderboards data87.4%
3Claude Sonnet 5.5· Steel.dev Agent Leaderboards data87%
4ERNIE 5.5 Flash Preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim84.9%
5Claude Haiku 5.5· Steel.dev Agent Leaderboards data81.5%
6Claude Opus 4.8· Steel.dev Agent Leaderboards data80.4%
7Claude Opus 4.7· Steel.dev Agent Leaderboards data77.7%
8Claude Opus 4.7web search/fetch + code · adaptive thinking · max effort · 1M token limit · compaction at 200K· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.77.7%
9Hy4 preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim77.2%
10Hy4 previewofficial model-card appendix· Reported by Tencent Hy· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.77.2%
11MiniMax M3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim
Values from 2 sources
73.2%
12MiniMax M3MiniMax Code Deep Research Skill · official rubrics · Opus 4.6 grader· Reported by MiniMax Research· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.73.2%
13Ling 3.0 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim70.4%
14Claude Fable 5· Steel.dev Agent Leaderboards data65.3%
15Claude Opus 4.6· Steel.dev Agent Leaderboards data59.8%
16gemini-deep-research· Steel.dev Agent Leaderboards data59%
17Claude Opus 4.5· Steel.dev Agent Leaderboards data46.7%

Data credits: Steel.dev Agent Leaderboards, MIT (licence) · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats