Skip to content

Benchmark

MRCR v2 · 1M Pointwise

Scores are percentages. Higher is better.

Models measured
10
Scores recorded
18
Sources
1
IQ weight
No weight published

Score distribution

43.3%
Mean
26.6%
Median
22.4%
Std dev
22.1%
10th percentile
74%
90th percentile
12.3%
Min
12.3%77.8%
The band covers the middle 80% of scores. The line marks the median.

Leaderboard

#ModelScore
1Claude Sonnet 4.6<1M API-reproducible subset · 29 problems · max effort· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.77.8%
2Claude Opus 4.6Anthropic Opus 4.6 launch page · adaptive thinking · max effort unless noted · 1M pointwise · 8-needle· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.76%
3GPT 5.5xhigh · 1M window · 512K–1M bin · 8-needle · mean sequence-matcher ratio· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.74%
4Claude Sonnet 4.6<1M API-reproducible subset · 54 problems · 64K extended thinking· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.71.3%
5Claude Sonnet 4.61M bin · full internal set · adaptive thinking · max effort · 5 trials· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.8%
6Claude Sonnet 4.61M bin · full internal set · 64K extended thinking · 5 trials· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.1%
7Muse Spark 1.1Meta Model API · xhigh · 1M window · 512K–1M bin · 8-needle · mean sequence-matcher ratio· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54.1%
8Gemini 3.6 FlashGemini 3.6 Flash Model Card · July 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54%
9Gemini 3.5 FlashGemini 3.5 Flash Model Card · May 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.6%
10Gemini 3.5 FlashGemini 3.6 Flash Model Card · July 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.6%
11Gemini 3 Pro PreviewGemini 3 Pro Model Card · results as of November 2025 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.3%
12Gemini 3.1 Pro Preview1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.3%
13Gemini 3.1 Pro PreviewGemini 3.5 Flash Model Card · May 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.3%
14Gemini 3.1 Pro PreviewGemini 3.6 Flash Model Card · July 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.3%
15Gemini 3.1 Pro Previewhigh · 1M window · 512K–1M bin · 8-needle · mean sequence-matcher ratio· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26.3%
16Gemini 3 Flash1M pointwise · 8-needle· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.22.1%
17Gemini 3 FlashGemini 3.5 Flash Model Card · May 2026 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.22.1%
18Gemini 3.1 Flash LiteGemini 3.1 Flash-Lite Model Card · high thinking · default sampling · single attempt/pass@1 · 1M pointwise· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.12.3%

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.