L2-Bench: leaderboard
Metric: Rubric score (%; weighted share of task criteria passed, negative criteria as penalties and task scores clipped at 0, mean over 1,000 second-language education tasks and three runs each; per-criterion pass/fail verdicts from a reference-guided Claude Sonnet 4.6 judge; provider-default inference settings). Source: arxiv.org. Saturation forecast: Around December 2026. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 | 85.5 |
| 2 | GPT-5.4 | 84.1 |
| 3 | Gemini 3.1 Pro (Preview) | 83.4 |
| 4 | Gemini 3 Flash | 80.7 |
| 5 | DeepSeek V3.2 | 80.2 |
| 6 | Kimi K2.5 | 79.1 |
| 7 | Claude Haiku 4.5 | 78.8 |
| 8 | Qwen 3 32B | 65.8 |
| 9 | Magistral Small | 50.7 |
Interactive version: theaggregate.ai/benchmark?slug=l2-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.