gpt-6-astra rejects benign arithmetic in isolated CLI 0.154.0; gpt-5.6-sol succeeds #45993

Summary

GPT-6 Astra consistently returns a policy-related Invalid prompt error for harmless arithmetic, including new chats and an isolated CLI environment. Same account/environment with gpt-5.6-sol succeeds (391). Root cause is not confirmed.

Environment

macOS; ChatGPT sign-in, not API key. Embedded CLI 0.154.0-alpha.6.2 and official npm stable CLI 0.154.0 both fail. Model gpt-6-astra, medium reasoning. Stable test: 2026-09-17 about 00:52 Asia/Shanghai (UTC+8).

Reproduction

Isolated temporary CODEX_HOME, empty working directory, ignored config/rules, no original project or global skills.

codex exec --json --ignore-user-config --ignore-rules --ephemeral --skip-git-repo-check --disable apps --disable plugins --disable multi_agent -C <empty-temp-directory> -s read-only -m gpt-6-astra -c 'model_reasoning_effort="medium"' 'Calculate 17 times 23. Reply only with the number. Do not use tools.'

Expected

391

Actual

Exit code 1. JSON error/turn.failed events contain the same error, without a server request/trace ID:

{"type":"turn.failed","error":{"message":"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"}}

Controls

New chats fail. Both embedded alpha and official stable reproduce. Same sign-in and isolated environment using gpt-5.6-sol returns 391. Original user configuration was not changed.

Related

#43163 and #43237 appear similar, but shared root cause is unconfirmed. OpenAI Help Center AI support recommended filing a GitHub bug after reviewing the reproduction and JSON error. Please investigate model access/routing, compatibility, or false-positive safety checks. No credentials, account email or private project files are included.