Astra for Coding: Why Are We Doing This Again?

Astra, the new GPT-6 model, excels at image understanding, complex topics and generating extensive code, but in a 35-hour, 4-billion-token experiment it failed to produce usable software and offered little insight into better workflows. The model disproportionately relies on Python for tool-call operations, often inserting excessive or unconventional Python snippets even within TypeScript, suggesting a training bias that rewards long-horizon output without penalizing low-quality code.