How to fix “Cannot read property choices of undefined” in OpenAI node?

The error happens because the OpenAI node is outputting a different JSON shape than you expect. The old n8n OpenAI (Chat / Completion) nodes used data.choices, but the new n8n OpenAI node (OpenAI Assistant, Chat Model, etc.) uses data in a different structure, so choices doesn't exist. Fix it by checking the actual output of the node (in the "Execution Data" panel) and updating your expression to match the real path. In most cases, you need either {{ $json.data[0].text }} or {{ $json.choices[0].message.content }}, depending on which OpenAI node type you're using.

 

Why the error occurs

 

The message Cannot read property choices of undefined means:

• You wrote an expression like {{ $json.choices[0].text }} or {{ $json.data.choices }}
• But the actual JSON the OpenAI node produced does not contain choices at that path
• So n8n tries to read it and fails, because choices is undefined

The key detail: n8n doesn’t normalize OpenAI output. The shape depends on the model type you select inside the node.

 

The three real OpenAI output formats in n8n

 

OpenAI nodes in n8n produce one of these **actual** structures:

• Chat Model (gpt-4, gpt-3.5, etc.)
  Output contains choices, like:
  \`\`\`json { "id": "chatcmpl-123", "choices": [ { "message": { "role": "assistant", "content": "Hello!" } } ] } \`\`\`
• Completion (legacy text-davinci-003)
  Also uses choices but inside choices[0].text.
• OpenAI Assistants / Threads
  Does not output choices at all. Structure looks like:
  \`\`\`json { "data": [ { "id": "msg\_123", "content": [ { "type": "text", "text": { "value": "Hello!" } } ] } ] } \`\`\`

If you use expressions written for the first format while your node is returning the third format, you get the error.

 

How to fix it step-by-step

 

• Open your workflow, run the node once manually.
• Click the OpenAI node → Execution Data.
• Look at what the JSON actually contains. Ignore tutorials, ignore assumptions — trust the real output.
• Update your expression to match the real path.

 

Correct expression examples

 

If your node output shows "choices":

// Chat model
{{ $json.choices[0].message.content }}

// Completion (old)
{{ $json.choices[0].text }}

If your node output shows "data" array (Assistants API):

{{ $json.data[0].content[0].text.value }}

 

A safe pattern to avoid the error in production

 

You can guard against undefined by checking existence:

{{ $json.choices?.[0]?.message?.content || '' }}

This won’t fix wrong paths, but it prevents the workflow from crashing. It’s useful in production where unexpected model changes or partial errors happen.

 

The root cause to remember

 

Every time you change model type in the OpenAI node, the JSON structure may change. n8n does not automatically adapt your expressions. Always check the real output of the node and rewrite expressions accordingly. That is the reliable production fix for “Cannot read property choices of undefined”.