Ploy reports a production agent moved to GPT-5.6 ran 2.2x faster at 27% lower cost

Category: Agentic coding
Status: discussion
Sources: Ploy engineering post, HN discussion
Summary: Ploy reports migrating its production website-building agent from Claude Opus 4.8 to GPT-5.6 Sol and evaluating both against a fixture suite of hundreds of cases. Median wall-clock per build dropped from 8m00s to 3m42s (2.2x), cost from 3.06 to 2.22 USD (27%), and output tokens from 33.0K to 17.1K, with a small visual-quality gain. The switch required reworking the eval harness (tool-call budgets were Opus-specific), a "required but nullable" tool-schema transform because GPT-5.6 fills every parameter, a workspace-scoped prompt-cache rebuild reaching a 83.7% hit rate, and disabling server-side reasoning replay.
Why it matters: The report quantifies that a frontier-model swap is mostly harness and caching engineering rather than a drop-in change, and that provider caching models differ enough to erase the gains if not rebuilt.