Hey everyone— we’re the Deep Principle team behind Mira AI Scientist. We just released Hydrogen-1.0 , an open-weight model post-trained from GLM-5.3 for chemistry, materials science, and life sciences.
Here’s how it performed in our evaluations:
FrontierScience Research: 38.3% pass rate vs. 33.3% for GPT-6 Astra at max thinking—a 5 percentage point lead .
Estimated output-side cost on FSR: $0.25 per problem vs. $0.88 for Astra at max thinking—about 72% lower under our evaluation setup.
GPQA Diamond: 92.9% accuracy, compared with Kimi K3’s 93.5% .
On separate internal task sets, Hydrogen also scored higher than GPT-6 Astra at max thinking on scientific computing (+5.13 percentage points) and single-step forward reaction prediction (+4.76 points).…