Attached files

Applications send images, PDFs, audio and video inline in the request body, as content parts (image_url/file/input_audio in Chat Completions, input_image/input_file in Responses, image/document in Anthropic Messages, and video_url for video models on vLLM and other OpenAI-compatible servers). The gateway has no upload endpoint: a provider file_id only works with the one provider account it was uploaded to, which would break alias routing and fallback.

Bodies are limited to 16 MB (Gateway__MaxRequestBodyBytes, at most 100 MB); larger requests get 413 request_too_large. Base64 adds a third, so this fits about 12 MB of files, enough for photos and PDFs but only short video clips. Send longer video as a URL the model server can fetch, or raise the limit. Whether a model accepts a given kind of file is up to the provider; Claude models take images, PDFs and plain text but not audio or video.

Budget reservations count each image as 1 600 input tokens instead of by its base64 size (a 4 MB photo used to reserve about a million tokens and could hit 402 budget_exceeded on a small budget). Documents, audio and video are still estimated from their size, which over-reserves; the actual cost is settled from the provider's reported usage.

Each key has an Attached files setting (Keys > edit key > Access):

A refused request gets 400 attachment_not_allowed. The PII policy reads text, not file contents, so use Text only for keys whose users handle data that must never reach a provider. A chat front-end that extracts a file's text itself and sends it as an ordinary message is not stopped by this setting; the PII policy covers that text. Details: admin-api.md.

Chat Completions requests to Claude models now carry PDF and plain-text files as Anthropic document blocks (they used to be dropped silently). Parts Claude cannot take (audio, video, other file types, OpenAI file_id) return 400 unsupported_content instead of being left out of the prompt.