A 27B model trained specifically to replicate research beat Claude Opus 4.8 and GPT-5.5.

Faraday (arXiv 2608.13331) narrows to reproducing published research and reports beating frontier general models on held-out replication. The size is what matters: 27B parameters beating frontier-plus-scaffolding on a bounded verifiable agentic task is the strongest recent argument for task-specific post-training. Self-reported and single-source, so treat as directional.