[Windows][Codex Desktop] GPT-6 models (Astra/Sol) reject every prompt with invalid_prompt, GPT-5.6 works fine #50657

What version of the Codex App are you using (From “About Codex” dialog)?

26.930.31730 (bundled codex-cli 0.160.0)

What subscription do you have?

ChatGPT 20*pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Summary
Every prompt sent to any GPT-6 model (gpt-6-astra, gpt-6-sol) is immediately rejected with
invalid_prompt ("Invalid prompt: your prompt was flagged as potentially violating our usage
policy"), including trivially harmless prompts like "hello" or "dd". GPT-5.6 models work
normally on the same account, same machine, same session.

Error (full text)
Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try
again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting

Underlying SSE event (captured from the raw Responses stream):
{"type": "response.failed", "response": {"status": "failed",
"error": {"code": "invalid_prompt", "message": "Invalid prompt: your prompt was flagged as
potentially violating our usage policy..."}}}

Steps to reproduce

1. Open Codex Desktop (or bundled CLI codex exec).
2. Select any GPT-6 model (gpt-6-astra or gpt-6-sol).
3. Send any prompt, e.g. "hello".
4. The request fails immediately with invalid_prompt.

Controlled experiment (rules out local factors)
Same account token, same network path, same trivial prompt ("Say hi"), sent directly to
https://chatgpt.com/backend-api/codex/responses with stream=true:

So the flag is tied to account + GPT-6 model family on the server side — not to the prompt
content, local config, project context, memory, or network path.

Additional notes

• Subscription is active (Pro, valid through 2026-10-13); account is in good standing,
  GPT-5.6 models fully usable.
• Reproducible in both Codex Desktop and the bundled CLI (codex exec), fresh sessions,
  no project files, no custom instructions.
• Failing CLI session IDs: 01a10218-5aa2-7230-b9b4-ba409905d236,
  01a10217-6211-7431-94db-bbd7cdb0ed88, 01a10216-bca5-7793-8e61-14731e31ac09
• Failing response IDs: resp_0f5e5c1d1d082f1f016ac11796b04c87d0a6fff20dbf306692 (gpt-6-astra),
  resp_063d8db712017503016ac11798bf6887d0be9839a7ef2d3760 (gpt-6-sol)
• This looks identical to the reports in GPT 6 - Potentially violating our usage policy on any/all prompts #47721, [Windows][Codex Desktop] GPT-6 Astra rejects a fresh task with invalid_prompt #47551, [Codex Desktop] GPT-5.6 Sol and GPT-6 Astra reject harmless prompts with invalid_prompt #47041, [Windows Desktop] GPT-6 Astra and GPT-5.6 Sol reject harmless prompts with invalid_prompt while Codex Web works #46756, False positive policy flag on harmless prompts, automatic model downgrade #46753 — appears to
  be a server-side moderation false positive / safety-routing regression rolled out with GPT-6.

What steps can reproduce the bug?

Feedback ID: 01a10240-acf3-7ac2-a539-fff4ccd977f1

What is the expected behavior?

Harmless prompts (e.g. "hello") sent to GPT-6 models should complete normally, the same way they do on GPT-5.6. Moderation should only flag genuinely policy-violating content. If the account is actually restricted from using GPT-6, the client should surface that explicitly (e.g. "your account does not have access to this model") instead of returning an invalid_prompt moderation error that blames the prompt content and sends users to the prompting guide.

Additional information

Account: ChatGPT Pro, subscription active through 2026-10-13, account in good standing.
Egress region: Taiwan (HiNet AS3462), direct connection, no proxy involved in the failing requests.
The flag triggers on 100% of GPT-6 requests regardless of prompt content, reasoning effort (tested xhigh and low), client surface (Desktop and bundled CLI), or session freshness.
Happy to upload a minimal failing session via /feedback or provide more response IDs if the team needs them.