- Ploy's agent now uses GPT-5.6 Sol as the default model, replacing Claude Opus after it performed better in quality, speed (2.2x faster), and cost (27% cheaper).
- Migrating to GPT-5.6 required fixing evaluation harness biases, tool call schemas (to handle optional parameters), prompt caching (workspace-scoped keys), and reasoning replay (self-contained blobs).
- GPT-5.6 excels at clean, modern designs but can be generic without guidance; tool call verbosity initially caused empty file reads, resolved via schema transforms.
- Prompt caching differences between providers significantly impacted costs; OpenAI's model requires explicit keys per workspace for effective cache hits.
- The model shows promise with lean code output and improved visual scores, but configuration adjustments are crucial for fair performance and cost comparisons.