CorpusQA (1M tokens)
General QA · 2026-01-21
CorpusQA tests corpus-level analysis and reasoning: questions must be answered by aggregating evidence spread across a very large document collection placed in context. This row is the 1M-token setting reported in the DeepSeek-V4 report, scored as answer accuracy; the benchmark itself scales to 10M tokens. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.6 | 71.7 |
| DeepSeek-V4-Pro | 62.0 |
| DeepSeek-V4-Flash | 60.5 |
| Gemini 3.1 Pro Preview | 53.8 |