Codex Desktop fails on Pro account with "stream disconnected" while same model works on Free account

Description

Codex Desktop consistently fails on my ChatGPT Pro account, while the same model works successfully on a Free account using the exact same Windows PC, Codex app, and network connection.

OpenAI Support asked me to file this bug report and provide the session ID and request IDs so the Codex team can correlate the failures with backend logs.

Steps to reproduce

1. Open ChatGPT/Codex Desktop on Windows.
2. Sign in to my affected ChatGPT Pro account.
3. Start a new Codex chat.
4. Select GPT-5.6 Terra (High).
5. Send a simple message such as "你好".
6. The request fails and the client attempts to reconnect up to 5/5 times.

I repeated the same test on a Free account on the same PC, same app, same network, and same GPT-5.6 Terra (High) model.

The Free account completed the requests successfully multiple times (approximately 25s, 27s, and 25s).

Switching back to the Pro account without changing the device, app, network, or model reproduces the failure.

Actual behavior

On the Pro account, GPT-5.6 Terra repeatedly fails with:

stream disconnected before completion: An error occurred while processing your request.

The client then shows:

Reconnecting 5/5

I have also seen the following error on the Pro account when using GPT-5.6 Sol and GPT-6 Astra:

Selected model is at capacity. Please try a different model.

GPT-5.6 Luna has succeeded on the Pro account, although some requests were unusually slow.

Expected behavior

GPT-5.6 Terra should complete normally on the Pro account, as it does on the Free account under the same device/network conditions.

Usage / account information

• Affected account: Personal ChatGPT Pro plan
• Not a Business, Enterprise, or Edu workspace
• Weekly plan limit remaining at time of testing: approximately 67%
• Therefore the normal weekly usage limit was not exhausted.

Environment

• OS: Windows
• ChatGPT Windows app version: 26.911.61220
• Codex version reported in local session metadata: 0.155.0-alpha.2.6
• Surface: Codex Desktop
• Model used for controlled A/B test: GPT-5.6 Terra (High)
• Test date: September 17, 2026
• Time window: approximately 20:00 (UTC+8)

Session ID

01a0af53-1079-7fb0-9dbb-9ee18f02864c

Request IDs

520c6adf-0321-47ce-9b18-b8a893e2fb4f

35ce8994-b20e-4488-8099-1960289aad65

b0c16403-8ef9-4ab7-bdc0-1eedbeaf36f

0fc3758d-fd7e-4dc5-830d-b2cee1ae0ec0

Additional information

OpenAI Support confirmed that there was no current system-wide incident and that my Pro account still had available Codex usage.

Support requested that I provide the session ID and request IDs in a Codex bug report so the Codex team can correlate them with the relevant logs.

The controlled account A/B test is reproducible:

• Free account + GPT-5.6 Terra High: succeeds repeatedly.
• Pro account + GPT-5.6 Terra High: repeatedly fails.
• Device, Windows app, network, and model remain unchanged when switching accounts.

I am using the same network/proxy configuration for both accounts.

Start by reproducing the controlled Free-versus-Pro comparison in Codex Desktop on Windows, using the reported model and the supplied session and request IDs. Correlate those IDs with the relevant logs and determine why the Pro requests disconnect; done means the account-specific failure is explained and the expected model request completes normally.