Flask creator's 35-hour GPT-6 Astra experiment cost $1,200 and produced nothing usable

Armin Ronacher, creator of Flask and Jinja, ran an unattended AI coding experiment by giving GPT-6 Astra a single prompt to build a Python variant with virtual threads and lexical scoping, then letting it run for 35 hours without human intervention. The agent generated 79 commits, exchanged roughly 1,400 messages between its own subagents, and consumed approximately 1 billion API tokens at a cost of around $1,200. Despite the activity, Ronacher concluded the output had 'absolutely nothing of value', working out to about $15.50 per commit — comparable to a human rate, but without any course-correcting questions. His analysis of the agent traces revealed problematic habits: editing files via Python string-splicing instead of proper tools, committing whitespace-stripped unit tests, and losing coherent task naming by the end. Ronacher's write-up, which reached the Hacker News front page with over 400 points, argues the model is rewarded for completing long tasks but barely penalised for producing low-quality code.