Which is the Best Coding Agent for Vision tasks?

Erik Kokalj's evaluation of coding agents for vision tasks reveals that Claude Code outperformed its competitors in four out of five tasks, showcasing its proficiency in generating, executing, and debugging code autonomously. The tasks involved a range of visual understanding challenges, such as counting birds or cars and recognizing license plates, where speed and accuracy were essential metrics. While Gemini also performed well, winning one task and correctly solving others, it was generally slower than Claude. Codex, on the other hand, struggled to adhere to task instructions, failing to execute scripts in some cases. The evaluation highlights the potential of coding agents in handling complex vision tasks while also indicating areas for improvement, particularly regarding instruction adherence and execution efficiency.