Claude thinking changes break cache, Codex may not

If this holds up, cache behavior matters almost as much as the model setting itself. A reasoning-level toggle that preserves cache can make Codex look faster and cheaper, but it also makes apples-to-apples testing much harder. Anthropic documents that changing thinking parameters invalidates cache breakpoints, so Claude's behavior matches the published rules. If Codex keeps cache across reasoning changes, developers need to pin cache state during evals or results will drift. Reused context is useful for iterative coding, but it can hide whether a reasoning-level switch actually changes behavior. This is a reproducibility issue as much as a performance one: latency, cost, and output quality all become harder to compare cleanly.