AlpacaEval 2.0 — Length-controlled win-rate is a distinct official metric variant from raw win-rate; kept separate
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | DeepSeek-R1 | DeepSeek | 87.6 | 6 | 2026-06-04 |
| 2 | DeepSeek-V3 | DeepSeek | 70 | 5 | 2026-06-04 |
| 3 | OpenAI o1-mini | OpenAI | 57.8 | 5 | 2026-06-04 |
| 4 | Claude 3.5 Sonnet | Anthropic | 52 | 5 | 2026-06-04 |
| 5 | GPT-4o | OpenAI | 51.1 | 5 | 2026-06-04 |
| 6 | R1-Zero | DeepSeek | 24.7 | 1 | 2026-05-03 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.