CLBench Life - Public Information Fragments: leaderboard

CL-bench Life (2026) tests whether language models can learn from long, messy real-life context to solve human-written tasks, a task counting as solved only when every expert rubric item is met. This board is the task solving rate in percent on tasks over feeds, news snippets, search results and public records; higher is better.

Metric: Task solving rate (%) on the 45 public-information-fragment (feeds, news snippets, search results, public records) tasks of CL-bench Life; a task is solved only when the response satisfies every human-written binary rubric item, judged by GPT-5.1 at high reasoning effort from the response and rubrics alone; reasoning mode at high effort where the model supports it, mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4High20#135
2Claude Opus 4.6High14.8#113
3Seed 2.0 ProHigh14.1#222
4Hy3-previewHigh13.3#322
5Grok 4.20High12.6#301
6Gemini 3.1 Pro (Preview)High11.9#99
7Kimi K2.5High9.6#266
8DeepSeek V3.2Thinking8.1#419
9Qwen 3.5 PlusHigh7.4#242
10MiniMax-M2.5Thinking4.4#600

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

Interactive version: theaggregate.ai/benchmark?slug=clbench-life-public-information-fragments · How It Works · Data refreshed daily, snapshot 2026-11-21.