LLM providers often silently cap context windows far below advertised limits

A developer running hosted chat and coding agents on open-weight models found that most API providers serve only around 32K tokens of context, regardless of the much larger figures stated on model cards. The advertised context window reflects the model weights, but the actual limit is set by the infrastructure operator, who must balance memory costs against serving capacity. The critical issue is not the cap itself but that it is rarely documented and fails silently — requests are simply truncated from the front without any error or warning. This causes subtle but serious problems in agentic and RAG workflows, where lost context can make a model appear incompetent at long tasks when the real culprit is a hidden serving limit. The developer calls for providers to expose the effective context ceiling per request through model metadata or response headers, rather than forcing developers to reverse-engineer it from error messages.