OpenAI has introduced a new feature allowing ChatGPT users to upload audio files directly for transcription, summarization, or querying. This functionality is available to paying subscribers and workspaces, including Enterprise, but not to free users. OpenAI notes that availability might vary depending on the region, model, application version, and workspace settings. While the underlying transcription technology has been available through APIs like Whisper and GPT-4o Transcribe, this update makes the process more accessible, as users no longer need to use an API or a specific tool to perform the task. ChatGPT supports audio files in several formats, including MP3, WAV, FLAC, AAC, M4A, and OGG, with a maximum file size of 512 megabytes. This is roughly nine hours of audio in MP3 format at 128 kbit/s, compared to about fifty minutes for an uncompressed WAV file. However, OpenAI clarifies that this is a sending limit and does not guarantee full transcription. Long recordings may be split into sections if Data Analysis is enabled, and particularly long files might be cut off before processing is complete. Additionally, transcription quality can vary depending on the language used. Testing conducted in October 2026 with two French telephone interviews revealed mixed results. A one-hour and 30-minute MP3 file was correctly identified but not fully transcribed. The system suggested using a third-party service called Descript. When the user requested not to use external tools, the assistant indicated it lacked the necessary engine. A 13-minute interview with three participants was partially transcribed, requiring the user to split the recording into two segments and re-import them. Even after this, the transcription covered only half of the interview, requiring the user to manually check and correct the content before obtaining the full transcription in the correct order. The final exchanges still lacked precise timestamps. In contrast, a transcription of the same file using Plaud, a specialized transcription tool, was completed in a single pass without losing half of the conversation and in a much shorter time. Plaud made more errors on certain names but identified only "Speaker 1" and "Speaker 2" for three participants. Despite these limitations, it processed the 13-minute and 40-second interview in one go, without the need for multiple steps or corrections. OpenAI acknowledges that transcriptions may contain errors and that participant identification might be unreliable. Testing highlights the need for users to verify not only names and quotes but also whether the entire recording was processed. Extra care is advised with sensitive content. On personal accounts, users can choose to allow OpenAI to use new conversations to improve models, though this can be disabled in data controls. By default, OpenAI does not train its models on content from Business, Enterprise, and Edu spaces. Uploading an MP3 file and requesting a transcription, summary, or search remains a useful feature. However, testing shows that users should still carefully preserve the original recording and verify what ChatGPT claims to have processed.