Databricks switches to GLM, achieving 92% pass rate at 1/4th cost

philfung · x · 2026-08-22

Databricks has switched internal workloads from Opus to GLM, finding that GLM achieves 92% of the pass rate at just one-fourth of the cost. This evaluation was conducted on proprietary codebases rather than public benchmarks to avoid potential hill-climbing. Combined with smart routing techniques and a shift to more efficient models, these strategies have driven unit costs down by up to 90% in some scenarios.

Original post →

More from Companies & People

Companies & People channel →