Ploy, an automated marketing website builder, has migrated its production AI agent from Claude Opus 4.8 to OpenAI’s GPT-5.6 Sol following the latter's release on July 9, 2026. The transition yielded a 2.2x increase in speed and a 27% reduction in operating costs, with GPT-5.6 Sol becoming the default model across all Ploy workspaces. The agent builds, edits, and optimizes websites by planning pages, reading code, and generating assets, creating a high-performance requirement for any model.

To successfully integrate GPT-5.6, Ploy engineers had to adjust their infrastructure to accommodate model-specific behaviors. The team discovered that GPT-5.6 tends to provide values for all 25 tool parameters by default, often inventing placeholder values that caused the agent to read empty files. By implementing a schema transform at the provider boundary that maps unused parameters to null, the team eliminated these empty reads and reduced total tool calls by approximately 30%.

The migration also required a complete redesign of Ploy’s prompt caching architecture. Unlike Claude’s org-scoped caching, GPT-5.6 uses a workspace-scoped key system with a capacity of roughly 15 requests per minute per cache node. After reconfiguring their system to use workspace-specific cache keys and implementing layered breakpoints, Ploy increased their first-call cache hit rate from 0% to 83.7%. This adjustment lowered total uncached input tokens by 28%, proving the cost-efficiency of the new model.

Finally, the team addressed reasoning state issues where the model's reliance on server-side pointers caused intermittent failures. By setting the SDK's reasoning state to self-contained blobs, they stabilized mid-conversation performance. Despite minor differences in design output, where GPT-5.6 requires specific steering to avoid generic layouts, Ploy confirmed the model now consistently outperforms previous incumbents in visual design scores, reaching 0.970 compared to Claude Opus's 0.936.

Ploy, an automated marketing website builder, has migrated its production AI agent from Claude Opus 4.8 to OpenAI’s GPT-5.6 Sol following the latter's release on July 9, 2026. The transition yielded a 2.2x increase in speed and a 27% reduction in operating costs, with GPT-5.6 Sol becoming the default model across all Ploy workspaces. The agent builds, edits, and optimizes websites by planning pages, reading code, and generating assets, creating a high-performance requirement for any model.

To successfully integrate GPT-5.6, Ploy engineers had to adjust their infrastructure to accommodate model-specific behaviors. The team discovered that GPT-5.6 tends to provide values for all 25 tool parameters by default, often inventing placeholder values that caused the agent to read empty files. By implementing a schema transform at the provider boundary that maps unused parameters to null, the team eliminated these empty reads and reduced total tool calls by approximately 30%.

The migration also required a complete redesign of Ploy’s prompt caching architecture. Unlike Claude’s org-scoped caching, GPT-5.6 uses a workspace-scoped key system with a capacity of roughly 15 requests per minute per cache node. After reconfiguring their system to use workspace-specific cache keys and implementing layered breakpoints, Ploy increased their first-call cache hit rate from 0% to 83.7%. This adjustment lowered total uncached input tokens by 28%, proving the cost-efficiency of the new model.

Finally, the team addressed reasoning state issues where the model's reliance on server-side pointers caused intermittent failures. By setting the SDK's reasoning state to self-contained blobs, they stabilized mid-conversation performance. Despite minor differences in design output, where GPT-5.6 requires specific steering to avoid generic layouts, Ploy confirmed the model now consistently outperforms previous incumbents in visual design scores, reaching 0.970 compared to Claude Opus's 0.936.