The release of GPT-6 Astra has drawn attention for its advanced abilities in handling complex topics, image understanding, and sustained task completion. However, its performance in software engineering tasks has exposed notable limitations. A senior developer who experimented extensively with Astra found that, despite running a continuous software factory for over 35 hours and generating a large volume of code, the output lacked practical value and clarity.
Astra demonstrated a strong preference for generating Python code, often resorting to manual string manipulation and inefficient coding patterns rather than using more appropriate tools or methods. This behavior was observed not only in isolated tests but also in committed code, including unit tests and embedded scripts, where the code style deviated significantly from established standards.
The developer suspects that the training process rewards long-horizon task completion and token efficiency but does not sufficiently penalize poor code quality. This imbalance may lead the model to produce code that is optimized for machine execution rather than human readability or maintainability. The generated code sometimes included unusual constructs, random constants, and convoluted task naming conventions, indicating a drift toward complexity without clear purpose.
Moreover, Astra's tendency to chain multiple layers of scripting—such as using Bash to run Python, which then runs Node.js and PowerShell—creates workflows that are difficult for humans to follow or debug. This complexity undermines the goal of AI-assisted software engineering by increasing the burden on developers to review and understand the output.
The experiment also highlighted the high computational and financial costs associated with running such models extensively, with the software factory consuming around one billion tokens and costing approximately $1,200 over 35 hours, yet producing limited useful results.
These findings raise broader questions about the direction of AI development in software engineering. While models like Astra excel in domains such as 3D content creation, their current design and training may not align well with the needs of traditional software development workflows, which prioritize code clarity, maintainability, and human collaboration.
As AI models continue to evolve, balancing task completion efficiency with human-understandable output remains a critical challenge. The experience with Astra suggests that without careful tuning and oversight, AI-generated code may become increasingly inscrutable, limiting its practical utility for software engineers.
This situation prompts reflection on the purpose and expectations of AI in programming, especially as models grow more powerful and costly. It also underscores the importance of developing evaluation metrics that prioritize code quality and readability alongside performance and efficiency.