n8n OpenAI Chat Model nodeConfigure OpenAI Chat Model in n8n.

Chooses which OpenAI model generates the completion. n8n loads the list dynamically from OpenAI when you open the dropdown, so what you see is what your account is allowed to call.

• Model: a dropdown filled at runtime from your account. A model you expected and cannot find is an account access question, not a node bug.

Overrides the default base URL for the API. Left empty, the node calls OpenAI at its normal address; filled in, every request from this node goes to the address you type.

• Base URL: a full endpoint address, applied to this node only and not to other OpenAI nodes in the same workflow.

Controls how much the model repeats itself. Positive values penalize new tokens based on how often they already appear in the text, which lowers the odds of the model repeating the same line verbatim.

• Frequency Penalty: a number. Higher reduces repetition; the further you push it, the more the wording drifts away from your prompt.

Caps the length of what comes back. A token is roughly a word fragment, and this sets how many the completion may use before it stops.

• Maximum Number of Tokens: a number. Most models carry a context length of 2048 tokens, so the value has to leave room for the prompt as well as the answer.

Decides whether the answer comes back as prose or as a JSON document. It is the simple, two-choice version of the setting, used on the Chat Completions path.

• Text (text): a regular text response, the default behavior.
• JSON (json_object): enables JSON mode, which should guarantee the message the model generates is valid JSON.

The richer format control available on the Responses API path, where you describe the shape of the answer instead of just asking for JSON.

• Type: text, json_schema, recommended, or json_object.
• Name: the name of the format. Letters, digits, underscores and dashes only, 64 characters at most.
• Schema: the JSON schema itself, with Description to explain it and Strict to force every answer to match it.
• Verbosity: low, medium or high.

Pushes the model toward new ground. Positive values penalize tokens based on whether they have appeared in the text at all, which raises the chance of the model talking about new topics.

• Presence Penalty: a number. Higher opens up new subjects; too high and the answer wanders off the question you asked.

Sets how random the sampling process is. Lower it and completions get less random; as it approaches zero the model turns deterministic and repetitive.

• Sampling Temperature: a number. Higher creates more diverse sampling but increases the risk of hallucinations.

Says how much thinking the model should spend before answering. More reasoning tokens means a more complete answer and a slower, heavier request.

• Low (low): favors speed and economical token usage.
• Medium (medium): balance between speed and reasoning accuracy.
• High (high): favors more complete reasoning at the cost of more tokens generated and slower responses.

Sets the maximum amount of time a single request is allowed to take, in milliseconds. Past that, the request is cut and the node reports an error instead of hanging.

• Timeout: a number in milliseconds, so 2000 means two seconds.

How many times the node tries again after a failed request before it gives up and fails the execution.

• Max Retries: a number. Each retry replays the same request, so a long batch takes proportionally longer when the provider is having a bad day.

Controls diversity through nucleus sampling: a value of 0.5 means half of all likelihood-weighted options are considered, and lower values simply ignore the less probable ones.

• Top P: a number. The general recommendation is to alter this or Sampling Temperature, not both at once.

Attaches the response to a conversation held on the provider side. Input and output items from the response are added to that conversation automatically once it completes.

• Conversation ID: a string, usually an expression such as {{ $json.conversation }} carried by the item, so one thread equals one conversation.

A key OpenAI uses to cache responses for similar requests, so repeated calls that share the key get better cache hit rates.

• Prompt Cache Key: a string you choose. Requests that should share a cache get the same value, requests that should not get different ones.

A stable identifier that helps OpenAI detect users of your application who may be violating its usage policies.

• Safety Identifier: a string that uniquely identifies each user, stable across their requests rather than regenerated on every run.

Picks the service tier the request runs under, which is how you tell the provider that this particular call matters more, or less, than the rest of your traffic.

• Auto (auto) and Default (default): the standard routes.
• Flex (flex): the tier to pick for work that can wait, typically a batch nobody is watching.
• Priority (priority): the one for work that cannot.

Attaches structured information to the object, useful for storing extra context and for querying objects later through the API or the dashboard.

• Metadata: a JSON set of up to 16 key-value pairs. Keys are strings of 64 characters at most, values strings of 512 at most.

Returns the most likely tokens at each position of the answer, each with its log probability. It turns the answer into something you can inspect rather than just read.

• Top Logprobs: an integer between 0 and 20, the number of candidate tokens reported per position.

Uses a reusable prompt template configured in the OpenAI dashboard instead of a prompt written inside the workflow.

• Prompt ID: the unique identifier of the template to use.
• Version: an optional version of that template, so a change in the dashboard does not reach production before you ask.
• Variables: the JSON values substituted into the template, often built from {{ $json }} fields of the incoming item.