ResearchBenchmarksLLM Pretraining on 500-shard corpus (val)Follow0.758Validation BPBOpus0.749520.806760.8640.92124Mar 31, 2026Evaluation ResultsMethodMethodLinksValidation BPBOpusvs. baseline=−22%, Cos...vs. baseline=−22%, Cost=$2002026.030.758Sonnetvs. baseline=−10%, Cos...vs. baseline=−10%, Cost=$1202026.030.869GPT-5.2Cost=$150Cost=$1502026.030.97