GPT-6.1 Sol rejects benign prompts as invalid_prompt while GPT-5.6 succeeds #52450

Summary

GPT-6-family requests in Codex Desktop consistently fail before generation with:

Invalid prompt: your prompt was flagged as potentially violating our usage policy.

The same error occurs for the benign prompt:

Reply with exactly: OK

Under the same ChatGPT Plus personal account, gpt-5.6-sol succeeds.

Reproduction

1. Open Codex Desktop on Windows.
2. Select GPT-6.1-Sol.
3. Use an empty ASCII-only local directory.
4. Send Reply with exactly: OK.
5. The request fails with invalid_prompt.

The issue was also reproduced from the CLI with user config ignored and plugins, apps, remote plugins, browser use, tool suggestions, skill search, and memories disabled.

Expected behavior

GPT-6.1-Sol should answer OK, as GPT-5.6-Sol does under the same account and environment.

Actual behavior

GPT-6.1-Sol rejects the benign prompt before generation as a potential policy violation.

Environment

• ChatGPT plan: Plus
• Workspace: Personal (not managed)
• OS: Windows 11, zh-CN
• Codex Desktop: 26.1002.7124.0
• Bundled Codex CLI: 0.162.0-alpha.2
• Also reproduced with CLI: 0.160.1
• Failure time: 2026-10-09 around 15:25 Asia/Shanghai (UTC+8)
• GPT-6.1 is visible/selectable in the model picker

Session IDs

• Failed GPT-6.1 minimal reproduction: 01a11f8d-98e5-7861-afd3-c8feebd123f5
• Successful GPT-5.6 control: 01a11f8e-0d52-7fd1-a56b-7a29aca04087
• Saved GPT-6.1 TUI reproduction: 01a11f9b-f7ef-7a73-8250-ada6c6c676a2

Diagnostics

codex doctor reports:

• ChatGPT authentication configured and valid
• Local configuration and databases healthy
• Responses WebSocket TLS handshake failure / routing discovery timeout
• ChatGPT inference endpoint timeout during doctor probe
• Desktop update/runtime CDN HTTP 403 or timeout
• Proxy environment variables point to 127.0.0.1:7890

I tested all seven dedicated ChatGPT proxy routes plus a general Japan route. The GPT-6 invalid_prompt behavior did not change. An empty directory, Chinese/English prompts, clean CODEX_HOME, stable/bundled CLI versions, and disabled local skills/plugins were also tested.

Feedback upload failure

I followed official support guidance and ran /feedback in the saved GPT-6.1 session. I selected safety check, included reproduction details, and approved upload of logs plus codex-doctor-report.json. The client returned:

Failed to upload feedback: feedback/upload failed in TUI

Therefore no feedback session ID was generated.

Request

Please investigate whether this is:

1. a false-positive safety decision for GPT-6 routes,
2. a GPT-6 entitlement/account-routing issue masked as invalid_prompt, or
3. a network-routing failure specific to GPT-6.