Responses API with code_interpreter returns empty output_text in v2.20.0 (regression from v2.19.0)

[acoustesh]

Bug report

Describe the bug

When using the Responses API with code_interpreter tool and file uploads, response.output_text is always empty in v2.20.0, even though the response status is "completed" and the code_interpreter successfully executes code. The same code works correctly on v2.19.0 and v2.18.0.

To reproduce

Minimal reproduction script:

import json
import os
from openai import OpenAI

client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

# Upload a tiny JSON file
data = {"items": [{"name": "A", "value": 10}, {"name": "B", "value": 20}]}
uploaded = client.files.create(
    file=("test_data.json", json.dumps(data).encode()),
    purpose="assistants",
)

try:
    response = client.responses.create(
        model="o4-mini",
        reasoning={"effort": "low"},
        tools=[
            {
                "type": "code_interpreter",
                "container": {"type": "auto", "file_ids": [uploaded.id]},
            },
        ],
        instructions="You are a helpful data analyst. Always reply with a final text answer.",
        input="Read test_data.json and return the sum of the 'value' column. Reply with just the number.",
        max_output_tokens=256,
        timeout=120,
    )

    print(f"Status: {response.status}")
    print(f"output_text: {response.output_text!r}")
    print(f"Output items: {len(response.output)}")
    for i, item in enumerate(response.output):
        print(f"  [{i}] type={item.type}")
finally:
    client.files.delete(uploaded.id)

Expected behavior

response.output_text should contain the model's text answer (e.g. "30"), as it does on v2.19.0.

Actual behavior (v2.20.0)

• response.status → "completed"
• response.output_text → "" (empty string)
• The code_interpreter_call output item shows status="completed" but outputs=null
• No message output item is present in response.output

Typical output on v2.20.0:

Status: completed
output_text: ''
Output items: 2
  [0] type=reasoning
  [1] type=code_interpreter_call  (status=completed, outputs=null)

Working behavior (v2.19.0 and v2.18.0)

Status: completed
output_text: '30'
Output items: 6
  [0] type=reasoning
  [1] type=code_interpreter_call  (status=completed)
  [2] type=reasoning
  [3] type=code_interpreter_call  (status=completed)
  [4] type=reasoning
  [5] type=message  (status=completed)

Version comparison matrix

openai version	o4-mini + effort=low	gpt-4.1-mini (no reasoning)
2.18.0	✅ PASS	✅ PASS
2.19.0	✅ PASS	✅ PASS
2.20.0	❌ FAIL (empty output_text)	❌ FAIL (empty output_text)

Environment

• OS: Ubuntu Linux
• Python: 3.12
• openai (broken): 2.20.0
• openai (working): 2.19.0, 2.18.0

Additional context

The issue affects both reasoning models (o4-mini) and non-reasoning models (gpt-4.1-mini), so it is not specific to reasoning effort settings. The regression appears to be in how the SDK handles the response deserialization — the API-side code_interpreter executes correctly, but the final message output item with the text response is missing from the parsed response on v2.20.0.

[acoustesh]

Same problem with newer models like "gpt-5.2"

[Rishi-Swami]

Thanks for the detailed report. I can confirm this seems to be a deserialization regression in the SDK, not a failure on the API side. I’m ready to work on a fix and add a regression test to restore the v2.19.0 behavior. Before I begin, I just want to check: is anyone already working on this, or is it okay for me to take it?

[desiorac]

Empty output_text regressions are dangerous in regulated AI systems. When a code_interpreter returns empty output but the call appears successful, your system logs show success, but the audit trail is incomplete.

For compliance: you need to distinguish between:

• "Code executed, output is empty" (valid, log it)
• "Code executed, output lost" (invalid, audit trail is broken)
• "Code failed, but not reported" (critical compliance failure)

This is why regulated AI systems need structured, machine-readable outputs—so you can detect and log failures reliably.

If you're building compliance-critical systems on top of OpenAI APIs, test against EU AI Act requirements: https://arkforge.fr/mcp (free, open-source)