n8n OpenAI nodeConfigure OpenAI in n8n.

Sends one or more messages to a model through the Responses API and returns its answer.

• Model: required, picked from the list or given by expression, such as {{ $json.model }}.
• Messages: each entry carries a Role (user, assistant or system) and a Prompt, and can hold an image or a file instead of text through Type.
• Conversation ID: attaches the response to an existing thread. It cannot be combined with Previous Response ID.
• Simplify Output: on, you get the answer alone; off, the raw payload.

Checks whether a piece of content complies with usage policies. The answer carries flagged, a list of categories and the matching category_scores.

• Text Input: the text to check, usually {{ $json.message }} straight from the trigger.
• Use Stable Model: on, the stable version of the model is used instead of the latest one, with accuracy that may be slightly lower.
• Simplify Output: returns a simplified response rather than the raw data.

Feeds one or several images to a model and answers a question about them, in text.

• Model: required, the model doing the reading.
• Input Type: url to pass URL(s) separated by commas, or base64 to read the binary from Input Data Field Name.
• Detail: low answers faster and consumes fewer tokens, high goes into detail, auto lets the model decide from the image size.
• Length of Description (Max Tokens): fewer tokens, shorter description. A token is the unit a model counts text in.

Creates an image from a text prompt and returns it as a binary file, or as a link when you ask for one.

• Model: dall-e-2, dall-e-3 or gpt-image-1.
• Prompt: up to 1000 characters for dall-e-2 and 4000 for dall-e-3.
• Number of Images, Resolution and Style: natural stays closer to a photograph, vivid pushes towards dramatic results.
• Respond with Image URL(s): on, you get links instead of binary files; off, Put Output in Field names the output field.

Reworks an image you supply, following the instruction in the prompt, instead of starting from a blank canvas.

• Prompt: required, up to 1000 characters for dall-e-2 and 32000 for gpt-image-1.
• Images: one or more binary fields, each a png, webp or jpg under 50MB, up to 16 images.
• Image Mask: a png whose fully transparent areas mark what to repaint, same dimensions as the image, under 4MB. With several images, it applies to the first one.
• Background and Input Fidelity: transparency and how closely the model matches the input style, both for gpt-image-1 only.

Reads a text out loud and returns the audio as a binary file.

• Model: tts-1 or tts-1-hd.
• Text Input: the text to voice, 4096 characters maximum.
• Voice: alloy, echo, fable, nova, onyx or shimmer.
• Response Format and Audio Speed: the container (mp3, opus, aac, flac) and the pace of the reading.

Takes an audio file from the incoming item and writes out what it says, in the language it was spoken.

• Input Data Field Name: defaults to data, and points at the binary property holding the file in flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav or webm.
• Language of the Audio File: an ISO-639-1 code, which improves accuracy and latency when you already know it.
• Output Randomness (Temperature): lower values keep the transcript closer to the audio.

Same input, different output: the audio comes back as English text, whatever language was spoken in it.

• Input Data Field Name: defaults to data, and accepts the same formats as a transcription, flac through webm.
• Output Randomness (Temperature): defaults to 1.0, and the range runs from 0.0 for deterministic to 1.0 for maximum randomness.

Removes a file from the server, permanently. Nothing else in the workflow gets touched.

• File: required, either picked by name from the dropdown or given as an ID, often {{ $json.id }} coming out of a listing step.

Returns the files that belong to the organization, one item per file.

• Purpose: narrows the result to one intent, assistants, fine-tune, vision or user_data. Left on any, everything comes back.

Pushes a file from the workflow to OpenAI so later operations can point at it by ID.

• Input Data Field Name: defaults to data, and names the binary property holding the file. Individual files go up to 512 MB, or 2 million tokens for Assistants.
• Purpose: assistants, fine-tune, vision or user_data. Fine-tuning only accepts .jsonl files.

Opens a thread that later responses can attach themselves to, and returns its ID.

• Messages: the opening content, each entry with a Role and a Prompt. A system message takes precedence over instructions given as user.
• Metadata: up to 16 key-value pairs, keys of 64 characters maximum, values of 512, for anything you need to search on later.

Fetches an existing conversation so the workflow can read what it holds.

• Conversation ID: required, in the conv_1234567890 form, typically stored earlier and pulled back with {{ $json.conversationId }}.

Deletes a conversation that is no longer needed. The thread and what it held go away.

• Conversation ID: required, the conv_1234567890 identifier of the thread to remove.

Changes the metadata carried by an existing conversation, without touching its messages.

• Conversation ID: required, the thread to update.
• Metadata: required here, a set of up to 16 key-value pairs, keys limited to 64 characters and values to 512.

Renders a short clip from a text prompt and puts the result in an output field as binary data.

• Model, Prompt and Seconds: all three required, the last one being the clip duration.
• Size: 720x1280, 1280x720, 1024x1792 or 1792x1024, the last two supported by Sora 2 Pro only.
• Wait Timeout: how long n8n waits for the render, in seconds, default 300. Output Field Name names where the file lands.