An open scientific AI model outperforms GPT-6 Astra on a research benchmark

With Hydrogen-1.0, Deep Principle specializes GLM-5.3 for scientific research. The open-weight model reports 38.3% on FrontierScience-Research, ahead of GPT-6 Astra.

What if the next breakthrough in scientific AI came not from an entirely new model, but from changing how machines learn to conduct research? With Hydrogen-1.0, Deep Principle is pursuing a different approach: training AI not only on scientific knowledge, but also on complete experimental workflows, from the initial hypothesis to the analysis of results, including mistakes and changes in strategy along the way. The company reports performance exceeding GPT-6 Astra on certain scientific evaluations and is releasing the model's weights so researchers can adapt it to their own work.

Introduced on November 2, 2026, by Chenru Duan, Hydrogen-1.0 builds on GLM-5.3, a model developed by Z.ai, with additional training focused on chemistry, materials science, and life sciences. Deep Principle reports a 38.3% pass rate on FrontierScience-Research, compared with 33.3% for GPT-6 Astra, alongside a 93% score on GPQA-Diamond. These figures come from the company's own evaluations and have not yet been independently validated. Nevertheless, they raise an interesting question: can an open-weight model, enriched with data from actual research processes, compete with the most advanced proprietary systems on specialized scientific problems? Learning Science From Experiments, Not Just Publications

Most language models acquire part of their scientific knowledge from research papers, books, databases, and other technical documents. These sources describe theories, methods, and results, but they don't necessarily capture the full sequence of decisions that led to a discovery.

A scientific paper generally presents the selected methodology and an organized set of conclusions. It often leaves out abandoned approaches, procedural errors, revised hypotheses, or experiments that produced no useful results. Yet these details play an important role in researchers' daily work, as they continually decide what to test, how to interpret an observation, and when to change direction.

Deep Principle wants to capture that dimension with Hydrogen-1.0. Its training draws on specialized scientific data, but also on the company's laboratory notebooks and documented discovery trajectories that record successive decisions made by researchers and AI agents.

The objective is to connect initial intentions with experimental protocols, observed results, and subsequent decisions. Rather than simply producing answers based on existing knowledge, the model is intended to learn how to organize experimental research and use previous findings to guide what happens next.

This approach doesn't guarantee that the model can formulate genuinely novel hypotheses or independently establish their validity. It does, however, attempt to incorporate something often missing from general-purpose training: how research evolves when confronted with imperfect, contradictory, or unexpected results. 38.3% on FrontierScience-Research, Ahead of GPT-6 Astra

Deep Principle accompanies the release with several evaluations designed to measure Hydrogen-1.0's scientific capabilities. The most widely highlighted result comes from FrontierScience-Research, a benchmark consisting of difficult scientific problems requiring multiple reasoning steps, particularly in physics, chemistry, and biology.

Hydrogen-1.0 achieves a reported 38.3% pass rate, compared with 33.3% for GPT-6 Astra in the company's evaluation. That represents a five-percentage-point lead. On GPQA-Diamond, a collection of advanced scientific questions, the model reaches 93% accuracy, which Deep Principle describes as competitive with leading open-weight models.

These results suggest that specialized training can improve the scientific capabilities of an already powerful general-purpose model. They do not, however, establish overall superiority over GPT-6 Astra. Performance depends on the questions selected, the amount of reasoning allowed, execution conditions, and the methods used to assess responses.

The financial comparison also requires clarification. Deep Principle presents average response costs across several benchmarks using publicly listed model prices. For Hydrogen-1.0, the company assumes pricing comparable to GLM-5.3. These estimates do not consistently include every expense, particularly input processing and evaluation costs, and should not be confused with the total cost of conducting scientific research.

The value of these measurements lies more in positioning the model than in establishing a definitive ranking. Independent replication under clearly documented conditions will be necessary to better understand the significance of the reported differences. Performance Varies Across Scientific Disciplines

Deep Principle also evaluated Hydrogen-1.0 on three categories of problems encountered in its research operations: scientific computing, chemical reaction prediction, and molecular structure elucidation using nuclear magnetic resonance, or NMR, data.

Against GPT-6 Astra operating at its maximum reasoning setting, Hydrogen-1.0 scores 76.9% on scientific computing, compared with 71.8% for its competitor. On predicting the products of chemical reactions, it reaches 90.5%, versus 85.7%.

The third category tells a different story. On NMR structure elucidation, Hydrogen-1.0 achieves 26.7%, while GPT-6 Astra reaches 30%. Deep Principle's model therefore does not consistently outperform its competitor, even within the scientific fields it was specifically trained to address.

This evaluation also illustrates the difficulty of certain scientific tasks. Determining a molecular structure from a spectrum requires interpreting multiple signals, comparing competing hypotheses, and checking their consistency with observed data. The relatively low scores are a reminder that strong scientific knowledge doesn't always translate into an ability to solve complex analytical problems.

These comparisons come from Deep Principle's internal evaluations, some of which use questions adapted from previous research. They provide insight