Hydrogen-1.0 Tops GPT-6 Astra With a 38.3% Research Pass Rate

Hydrogen-1.0, a research model developed by Deep Principle using GLM-5.3 as its base, recorded a 38.3% pass rate on OpenAI’s FrontierScience Research evaluation, compared with 33.3% for GPT-6 Astra. The figures are pass rates from Deep Principle’s results, rather than raw point scores.

The company trained the model on laboratory records describing experiment design, errors and changes made during research. Those records capture work that scientific papers typically leave out when they present final findings. The demonstrations put the model through tasks involving code, scientific software, lab protocols and materials screening.

In one demonstration, Hydrogen-1.0 used Python and GROMACS in a polyethylene melt simulation. It detected and repaired a topology error, then restarted the simulation. In a separate virtual lab rehearsal, it prepared a JSON protocol for a 12-vial Suzuki coupling study. The model caught an attempt to aspirate without a fresh pipette tip and revised the order of actions.

Deep Principle also ran five rounds of new-material screening, with each round informed by the previous evaluation. The company compared the models’ mean best-so-far HOMO–LUMO gaps: 4.318 eV for Hydrogen-1.0, 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. These are model-level results across the screening rounds, not a report that one final candidate surpassed the other models.

Hydrogen-1.0’s model weights are available under the MIT license. The full datasets and recipes used for training were not released with the weights.