The experiment involved reconstructing the Pomotroid v0.13.0 desktop timer using the gpt-6-astra model via three different interfaces: OpenAI's official API, OpenRouter, and TeamoRouter. The task required analyzing the application's source code, implementing core timer logic, and adding JSON import/export functionality within an isolated Linux Docker environment.
All three versions successfully passed nine automated tests, achieving a score of 70/70. The baseline timer logic accounted for 40 points, settings persistence for 10 points, and the import/export feature for 20 points. The remaining 30 points were reserved for manual evaluation.
Performance metrics varied across the platforms. The official OpenAI API completed the first stage in 6 minutes 23.1 seconds, OpenRouter in 6 minutes 49.1 seconds, and TeamoRouter in 8 minutes 46.6 seconds. Total API requests ranged from 27 to 33 per run.
Cost analysis showed significant differences. The official OpenAI API cost $5.82 for 29 requests, while OpenRouter cost $5.71 for 27 requests. TeamoRouter was the most cost-effective at $1.67 for 33 requests, though it required more time and tokens to complete the task.
