GPT 5.1 on the API (through Azure AI foundry) outputting what looks like reasoning tokens in the start of the final output (happened between tool calls).
Unfortunately the reasoning is mixed danish and english, so might be hard to understand for you, but they do seem like the “real” reasoning tokens and not a hallucinated version of it, especially since it aligns with tool calls. There is nothing in system prompt that indicates that it should do tool call preambles either. It does seem like it did some reasoning before the first ‘google’ tool calls, but was not part of the final completion tokens. But the next 2 sections are. Marked with red in the picture.
Very strange. Can’t recreate the bug.
