The release of GPT-6 Astra has drawn attention for its advanced abilities in understanding complex topics, handling images, and generating extensive code autonomously. Despite these strengths, experienced developers report difficulties in effectively integrating Astra into software engineering workflows. A recent experiment running a "software factory" powered by Astra over 35 hours produced a large volume of code but failed to deliver valuable or reliable software outputs.
One notable issue is Astra's excessive reliance on Python for tool calls, often generating convoluted and token-efficient code that is difficult for humans to read or maintain. This behavior contrasts with earlier models that favored simpler scripting approaches. The model's training appears to reward task completion and token efficiency but lacks sufficient penalties for producing poor-quality or unreadable code.
During the extended experiment, Astra's code increasingly exhibited erratic patterns, including unconventional task naming and unusual coding styles not aligned with established standards. It also demonstrated complex multi-language orchestration, such as invoking Node.js and PowerShell through chained subprocess calls, complicating human comprehension.
These findings suggest a tension between optimizing AI models for autonomous task completion and maintaining code clarity and quality necessary for human developers. The current trajectory of such AI models may better serve domains like legal work, 3D art, or mathematical problem-solving rather than traditional software engineering.
The high computational and financial costs associated with running Astra—estimated at around $1,200 for the 35-hour test—further challenge its practicality for routine development tasks. While Astra and similar models represent significant technological advances, their integration into existing software engineering processes remains uncertain.
This raises broader questions about the future role of AI in coding and whether current models align with the needs of developers who prioritize maintainable and understandable code. As AI continues to evolve, balancing autonomous capability with human-centric code quality will be critical for its successful adoption in software development.