MRCR (Multi-Round Co-reference Resolution) — 1M-context, explicit MMR metric; DeepSeek-V4 source, distinct from v2 variants.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6 | Anthropic | 92.9 | 4 | 2026-07-04 |
| 2 | DeepSeek-V4-Pro | DeepSeek | 83.5 | 9 | 2026-06-27 |
| 3 | DeepSeek-V4-Flash | DeepSeek | 78.7 | 4 | 2026-06-27 |
| 4 | Gemini 3.1 Pro | 76.3 | 4 | 2026-07-04 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.