Docbench - Study: leaderboard

Docbench in Gezel: fixed document-task adapters, parser/repair/retry behavior, native checks plus Claude Opus 5.5 rubric judge. Unusable outputs score zero. 32, 768 context except Apple 8, 192. Mostly Q4 is not affirmative precision evidence for each row; preserve explicit Q2/Q4 only. CLI/API labels are run configuration, not distinct public models. Native Apple Foundation Models row retained in pinned capture but not emitted until exact checkpoint identity is established. 102 public cases, 142 outputs; some models have three repetitions. Overall family weights25/25/20/10/15/5; 95% range is uncertainty, not another observation. Retain each native family grader mixture. Metric: Native task quality score (%). Source: https://bendyline.com/posts/docbench.html

Metric: Native task quality score (%). Source: bendyline.com. Saturation forecast: Around February 2027. 29 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 5.584#7
2Claude Sonnet 5.584#21
3GPT-6.1 Sol83#12
4GPT-5.6 Sol81#33
5Claude Haiku 4.579#580
6Gemma 4 31B78#385
7Claude Sonnet 4.677#173
8Qwen 3.6 35B A3B74#484
9Qwen 3.8 27B (Q4)74
10Qwen 3.8 27B (Q2)74
11Gemma 4 26B A4B73#581
12Ornith-1.5-9B73#843
13Gemma 4 12B72#931
14Qwen 3.5 9B71#822
15Gemma 4 E4B69#2142

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=docbench-study · How It Works · Data refreshed daily, snapshot 2026-11-17.