Hangzhou startup claims Hydrogen-1.0 beats GPT-6 Astra on research benchmark
Deep Principle's Hydrogen-1.0, built on GLM-5.3, reports 38.3% on OpenAI's FrontierScience-Research benchmark versus 33.3% for GPT-6 Astra.
SITUATION EXPLAINED: A Hangzhou startup says its open-weight AI model beats OpenAI's GPT-6 Astra at scientific research. • Hydrogen-1.0 is built on GLM-5.3, an open model from Chinese lab Z .ai • Deep Principle trained it further on its own lab records, including how its scientists actually worked through problems • FrontierScience-Research is a benchmark OpenAI built itself, with 60 multi-step research problems written by PhD scientists • On it, Deep Principle reports 38.3% vs 33.3% for GPT-6 Astra • It also reports 93% on GPQA Diamond, a set of PhD-level science questions written so the answers can't easily be looked up online • The model is priced at the same rates as GLM-5.3 for anyone using it through an API @sophiadew: "You actually wanna see all of the research notes, all the things that went wrong, everything that actually happened within the laboratory, and that's how you can train these models better for science."
This post quotes another post: x.com/1110539695719628800/status/211724929246966…