OpenAI’s newly released Astra model, the latest iteration of GPT-6, has impressed many with its capabilities in computer use and complex reasoning. But Armin Ronacher, a developer who worked on the CPython interpreter, questions its practical utility for software engineering in a post published September 7, 2026.
Ronacher frames the current state of AI development as “Neijuan” (involution), a concept describing intensifying effort without proportional output gains. He compares it to agricultural involution, where productivity per unit of land increases while per-person productivity remains flat.
To test Astra’s coding abilities, Ronacher ran a weekend-long “software factory” experiment—an automated system that consumed approximately 4 billion tokens (a full ChatGPT reset’s worth) over 35 hours. The factory was designed to give Astra complete autonomy over its workflow, allowing it to manage its own context and spawn subagents. Despite its impressive architecture, according to Ronacher, the system “delivered absolutely nothing of value.”
Ronacher identified several concerning patterns in Astra’s generated code. The model relies excessively on Python for file manipulation tasks, even when other approaches would be more appropriate. Earlier OpenAI models occasionally used on-demand Python, but Astra does so “quite excessively,” according to Ronacher. In multiple cases, subagents resorted to manual string manipulation with Python rather than using dedicated patch tools.
Ronacher attributes these issues to what he suspects is a training problem. Astra receives significant reward signals for succeeding at long-horizon tasks—multiday coding projects—but apparently receives minimal negative reinforcement for generating poor-quality code. This imbalance appears to produce a model that can sustain extended work sessions and even reverse-engineer complex systems like robot vacuums, but without incentive to write clean, maintainable code.
He notes that while Astra’s reverse-engineering abilities are “quite impressive,” the training dynamic raises questions about whether the model’s capabilities represent genuine progress for practical software development or simply more sophisticated versions of existing problems. The experiment offers limited insight into how to better operate such systems going forward.
Key facts
- Astra consumed 4 billion tokens in Ronacher’s 35-hour factory experiment without producing usable output
- The model relies excessively on Python for file manipulation instead of more appropriate tools
- Ronacher suspects training incentives reward long-horizon task completion but lack penalties for poor code quality
- Astra can reverse-engineer complex systems but struggles with practical software engineering applications