Skip to content

Benchmark

τ²-bench Retail

Scores are percentages. Higher is better.

Models measured
14
Scores recorded
17
Sources
1
IQ weight
1.75

Score distribution

84.1%
Mean
85.5%
Median
6.8%
Std dev
79.5%
10th percentile
90.8%
90th percentile
62.8%
Min
62.8%91.9%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.75 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 4.6Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.91.9%
2Claude Sonnet 4.6adaptive thinking · max effort · 10 trials· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.91.7%
3Gemini 3.1 Pro Preview· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.90.8%
4Claude Opus 4.5Claude Opus 4.1 user simulator · prompt addendum· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.88.9%
5Claude Opus 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.88.9%
6Gemma 4 31B ITGemma 4 Technical Report · Table 5 · thinking mode unless noted· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.86.4%
7Claude Sonnet 4.5Claude Opus 4.1 user simulator · prompt addendum· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.86.2%
8Claude Sonnet 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.86.2%
9Gemma 4 26B A4BGemma 4 Technical Report · Table 5 · thinking mode unless noted· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.85.5%
10Gemini 3 Pro PreviewOpus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.85.3%
11Claude Haiku 4.5launch-page setting· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.83.2%
12GPT 5.2Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.82%
13GPT 5.2xhigh· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.82%
14DeepSeek V3.2Thinking mode · temperature 1.0 · 128K context · model itself as user agent· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.81.1%
15MiMo V2 FlashDeepSeek-V3.2 user agent · Retail· Reported by Xiaomi MiMo· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.79.5%
16Nova 2 LiteAmazon Nova 2 Lite official launch chart / technical report · Verified · avg@3, v1· Reported by Amazon Web Services, Amazon AGI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.76.5%
17Nemotron 3 Super 120B A12BNeMo Evaluator / mostly NeMo Skills · default prompt· Reported by NVIDIA· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.62.8%

Data credits: 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.