Chat completions stream cut off without finich_reason (regression after update)

[matthijs166]

Confirm this is a Node library issue and not an underlying OpenAI API issue

• This is an issue with the Node library

Describe the bug

Its a bit hard to exactly pinpoint whats going on but we have an application running for some time now on version 6.9.1 of the sdk where every thing works as expected.

When we upgraded to version 6.16.0 the stream response was cut of mid generation...

To Reproduce

The following code works with SDK version 6.9.1 and cuts off half way in version 6.16.0 without finich reason.

Keep in mind that the input tokens are in total over 50k and we expect over 40K output tokens. (it takes some time to complete..)

// Setup of client
const client = new OpenAI({
	apiKey: key,
})

// Start completion
const response = client.chat.completions.create({
  messages: [
			{ role: 'system' as const, content: systemPrompt || 'You are a helpful assistant.' }, // dynamic system prompt
			{ role: 'user' as const, content: input }, // dynamic input
		],
  model: 'gpt-5.1-2025-11-13',
  stream: true,
  stream_options: {
	include_usage: true,
  },
  user: '123', // dynamic user id
  store: false,
  reasoning_effort: 'none',
  response_format: {
	type: 'json_schema',
	json_schema: {
		name: 'schema_name', // dynamic name
		strict: true,
		schema: z.toJSONSchema(zodSchema), // dynamic zod schema here
	},
   },
})

// Handeling incomming data
let fullResponse = ''
let finishReason: ChatCompletionChunk.Choice['finish_reason'] | null = null
let usage: OpenAI.Completions.CompletionUsage | null = null
for await (const chunk of response) {
	const choice = chunk.choices[0]
	if (!choice) continue

	if (choice.delta.content) {
		const newGeneratedToken = choice.delta.content
		fullResponse += newGeneratedToken
	}

	if (choice.finish_reason) {
		finishReason = choice.finish_reason
	}

	if (chunk.usage) {
		usage = chunk.usage
	}
}

// Handeling errors etc..
if (!finishReason) {
	throw new Error('OpenAI response has no finish reason')
}

// Do some more work

Code snippets

OS

macOS - 26.2

Node version

node v22.20.0

Library version

openai 6.16.0

[nomiveritas]

Set timeout: 0—this is crucial. It bypasses the SDK’s internal 10-minute timer that kills long processes.

Why it’s mandatory for 40k+ tokens:
-SDK bug workaround: v6.16.0 mishandles default timeouts on long streams.
Gives the model enough time:** 40k tokens can take 15–25 mins, longer than default allows.

typescript
const client = new OpenAI({
apiKey: key,
timeout: 0, // fixes the SDK timeout bug
});
With this, your for await loop will reach the last chunk, return finish_reason, and stop throwing errors.

[matthijs166]

@nomiveritas

Set timeout: 0—this is crucial. It bypasses the SDK’s internal 10-minute timer that kills long processes.

Isn’t this only necessary when streaming is off?
With streaming enabled, bytes are sent regularly.

I will test soon if it works!

[nomiveritas]

The timeout: 0 setting is required only for long-running processes without streaming, as the SDK waits for the full response and the default 10-minute timer may terminate the request. With streaming enabled, continuous data transmission keeps the connection active, so adjusting the timeout is unnecessary.
While streaming often keeps the connection alive, it does not fully eliminate SDK or infrastructure timeouts in long-running requests. For safety, timeout: 0 is recommended for long processes, regardless of streaming. The default 10-minute SDK timer can otherwise terminate the request before finish_reason is emitted. Testing remains essential.
Respectfully,
Veritas

[matthijs166]

The request did timeout immediately in version 6.9.1 "Request timed out."

We have an new update cycle next month, will try in newer versions