Repository navigation
GPT-6 Astra rejects harmless prompts and falls back to GPT-5.6 Sol #44700
Description
What version of the Codex App are you using (From “About Codex” dialog)?
26.903.8094.0
What subscription do you have?
Plus
What platform is your computer?
Microsoft Windows 11 Pro 10.0.26200 x64
What issue are you seeing?
GPT-6 Astra rejects harmless prompts such as “你好” or “Hello” with the following error:
Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt.
Codex then automatically changes the model from GPT-6 Astra to GPT-5.6 Sol. GPT-5.6 Sol works normally.
The issue also occurs in a completely new task with no previous conversation context.
Local logs around the reproduction contain:
- DeviceCheck attestation unavailable (errorCode=1)
- Received turn/started for unknown conversation
What steps can reproduce the bug?
- Open the Windows Codex desktop app.
- Create a new empty task.
- Select GPT-6 Astra.
- Send “你好” or “Hello”.
- The prompt is rejected as a usage-policy violation.
- Codex automatically falls back to GPT-5.6 Sol.
Reproduction rate: 100%.
Expected behavior:
GPT-6 Astra should respond normally to harmless greetings.[
](url)
What is the expected behavior?
No response
Additional information
No response
Activity
github-actions commentedon Sep 11, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
freanocank commentedon Sep 12, 2026
I can reproduce this on macOS with the same Codex account:
Invalid prompt: your prompt was flagged as potentially violating our usage policy.codex_error_info: "other"and no assistant/output item in the local session record, while the Sol control run produces a normal assistant message.This looks like an Astra-specific failure before model output (possibly session/turn initialization or gateway validation), with a misleading generic
invalid_promptmessage.freanocank commentedon Sep 12, 2026
Follow-up diagnostic: with the current Codex App 26.908.40834 / bundled CLI 0.154.0-alpha.6.2, a clean temporary CODEX_HOME, an empty workdir, and
--ignore-user-config, GPT-6 Astra still fails onhiwhile GPT-5.6 Sol succeeds. The authenticated direct backend request also reachesresponse.in_progressand then returns SSEinvalid_prompt/response.failedfor Astra, while the identical Sol request completes. This is not a local project/configuration or capacity issue. Detailed transport evidence and request IDs are in my follow-up on #43237: #43237 (comment)