Research
Experiments we run on our own factory
H2M builds software with coding agents. Before a method goes into client work it runs here as an experiment: a fixed question, a frozen spec, a hidden test suite, and every number published, including the ones that did not work.
Newest first
All research
How GLM-5.3 and GLM-5.3-Flash built the same clinic website in 15 hours of run time
GLM-5.3-Flash and GLM-5.3 each built the same clinic website alone from one frozen backlog: 100% and 97.8% of hidden tests passed, for $11.49 and $136.73.