Reasoning / Industry term
Hidden chain-of-thought
Intermediate reasoning tokens or traces that a model provider does not expose directly to the user.
Some reasoning models produce an internal token sequence before the visible answer. The product may return only the answer or a separate summary rather than the raw trace. That trace is generated computation, but its language is not guaranteed to be a complete or faithful explanation of every factor behind the result. Providers may hide it for monitorability, safety, product, or intellectual-property reasons.
Builder example
Hidden reasoning can affect latency, usage, and cost, although billing fields differ by provider and model. Because applications cannot depend on the raw trace, they should expose observable evidence such as tool calls, source support, tests, and decision records instead of treating a generated explanation as an audit log.
Common confusion: A hidden chain of thought is neither direct access to model internals nor a guaranteed faithful account of how the answer arose. It is a generated trace whose relationship to the model's computation is an active research question.