A paying customer reported that generation kept failing whenever he selected our most capable model. Two earlier fixes (a deprecated API parameter, then a rejected temperature setting) each resolved a real error — but he was still hitting failures after both.
The actual cause was different in kind, not just detail: GPT-5.6 is a reasoning model. Before it writes anything a user sees, it spends part of its own response budget on internal reasoning — tokens the API bills for but the user never reads. At the token ceiling we had set (sized correctly for a non-reasoning model), that invisible reasoning could consume the entire budget on its own, leaving nothing for the actual answer. The result wasn't an error — it was a silent, valid, empty response.
The fix was two-part: cap reasoning effort at a lower setting, and raise the token ceiling specifically for reasoning models — comfortably above what the provider's own documentation recommends reserving for reasoning plus output combined.
The broader lesson: a “newer, more capable” model isn't a drop-in replacement for an older one from the same vendor. Reasoning models have a genuinely different resource-consumption shape, and code written for a non-reasoning model can fail in ways that look like a bug in your own integration — because in a real sense, it is one, just not the one you'd assume.