Field Notes

Question Everything runs several AI models in production, every day. This is the real, dated record of what we've actually found running them — not what a vendor's docs say should happen, what we've genuinely observed happen, with the fix.

Every entry here is a real incident. Nothing is hypothetical.

GPT-5.6
A model can burn its entire answer budget “thinking” and return nothing
GPT-5.6 is a reasoning model — it spends part of its response budget on invisible reasoning before writing anything visible. At the token budget we'd set (correct for a non-reasoning model), that reasoning could consume the whole budget and return an empty response, with no error to explain why.
GPT-5.6
A required parameter changed name, and every request to the new model silently failed
OpenAI's newer models reject the max_tokens parameter outright, requiring max_completion_tokens instead — a breaking change that fails every request until you know to look for it.
GPT-5.6
The same model family rejected a setting we'd always been able to send
After fixing a deprecated-parameter error, the same model kept failing — a second, unrelated restriction: newer reasoning models only accept their own default temperature, not a custom value.