Anthropic introduced Agent4Sci, a long-horizon benchmark framework for scientific machine learning tasks covering 11 scientific domains, 20 benchmarks and 25 tasks. Experiments showed four frontier agents improved all 20 benchmarks collectively, with Claude Opus achieving 10.4% average relative improvement, rising to 20.1% with added plugin skills. The work was accepted into the NeurIPS 2026 Evaluation & Datasets track.