SAHARA - Reading Comprehension: leaderboard
Reading comprehension on Belebele multiple-choice passages in African languages, from the SAHARA African-language LLM leaderboard (UBC-NLP, ACL 2025). The task score is the mean over the languages the task covers; Accuracy on a 0-100 scale, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around 2029. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 58.89 | #71 |
| 2 | Gemini 2.5 Pro | 55.92 | #159 |
| 3 | GPT-5 | 51.01 | #105 |
| 4 | GPT-5.6 Luna | 51 | #79 |
| 5 | GPT-5.1 (2025-11-13) | 50.24 | #107 |
| 6 | Gemini 2.5 Flash | 48.01 | #255 |
| 7 | Claude Sonnet 4 | 42.33 | #207 |
| 8 | GPT-4.1 | 39.7 | #258 |
| 9 | Qwen 3.6 35B A3B | 39.44 | #238 |
| 10 | Qwen 3.5 27B | 37.57 | #192 |
| 11 | Qwen 3.6 27B | 36.96 | #182 |
| 12 | Claude Sonnet 4 (20250514) | 36.6 | #223 |
| 13 | Llama 3.3 70B Instruct | 35.92 | #541 |
| 14 | DeepSeek R1 Distill Llama 70B | 33.72 | #614 |
| 15 | DeepSeek R1 Distill Qwen 32B | 33.64 | #662 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sahara-reading-comprehension · How It Works · Data refreshed daily, snapshot 2026-11-05.