Windsurf Flex tier default model swapped to a smaller variant — quality drop measurable on NerfBench
Codeium's Windsurf editor silently routed Flex-tier requests to a smaller/cheaper underlying model. The model identifier in the model picker remained the same, but per-prompt performance dropped measurably.
What changed
Panel verdict· panel not assembled
The panel is convened only for receipts at major severity or above. INFO and minor receipts are filed for record without LLM verdicts.
Response· vendor statement · uncontested
UNCONTESTEDWindsurf has not publicly addressed this change.65 days since filing · response window has closed · this receipt is marked uncontested.
Predicted impact · est.
low confidenceHeuristic estimate. Quality regression hits everyone using the model. How we estimate ↗
Context · narrative
Discovered when independent benchmarks (including NerfBench) showed a sudden ~12% pass-rate drop on identical prompts. Windsurf acknowledged the swap in a community thread but did not update marketing claims about model quality.