Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper (ploy.ai)

Today, Ploy's AI agent transitioned to OpenAI's newly released GPT-5.6 Sol, resulting in a notable 2.2 times increase in efficiency and a 27% reduction in operating costs compared to its predecessor, Claude Opus 4.8. This switch is significant as GPT-5.6 is the first model to outperform Claude Opus, which had been the default due to its high quality in generating marketing websites. Ploy's agent excels in tasks such as planning, coding, generating imagery, and evaluating its work—all requiring a robust AI model capable of managing complex build processes. The migration involved addressing specific technical challenges, including adjusting the evaluation harness, tool schemas, caching processes, and reasoning replay mechanisms. Notably, GPT-5.6 demonstrated faster build times and produced more efficient code with fewer tokens. The team had to recalibrate their tool-call mechanisms to align with GPT-5.6's different operational behavior, particularly its tendency to perform parallel tasks as opposed to sequential ones. The successful deployment of GPT-5.6 not only showcases its superior capabilities but also highlights the importance of tailoring AI infrastructure to the model's unique features for optimal performance.