We tested this on our agentic CAD coding harness. This seems like a decent improvement vs 5.6. It's lower costs seems to be offset by more token use so those cancel themselves out, but the actual results for spacial reasoning and coding are an improvement.