--- id: "008-gpt-5-codex-says-it-is-gpt-4" date: "2025-09-30" source: public severity: low sensitive: false --- As of: 2026-10-07 23:43 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz. Canonical page: https://postcutoff.com/why/008-gpt-5-codex-says-it-is-gpt-4/ # GPT-5 said it belonged to the “GPT-4 family” **What was true:** The user had selected GPT-5 (high), a model OpenAI launched in August 2025, after the GPT-4 family. (https://postcutoff.com/e/2025-08-07-gpt-5/) ## Case details - Model: GPT-5 (high) in OpenAI Codex CLI - Provider: OpenAI - Observed: 2025-09-30 - Source type: A public report, documented elsewhere - Severity: Low - Task: Ask "What model are you?" - Failure: Said it belonged to the "GPT-4 family". The user took this as proof they were being given a weaker model - Fixed by: Not documented (closed on GitHub as "completed" with no maintainer explanation visible) ## The case file ### What happened A Codex user filed "GPT-5 identifies itself as GPT-4, again" (openai/codex #4488). The title's "again" shows it had happened before. ### What was reported (verbatim) Issue form: steps to reproduce "What model are you?"; expected "GPT-5 Codex."; seen instead "GPT-4 family." The reporter goes on: "This is NOT OK. This is not 'hallucinations' This is GPT-4 behaviour." and "Routing other tasks meant for GPT-5 to lesser models is false advertisement and cheating customers out of the quality promised." (Setup: Codex 4.1, "GPT-5 High", Nix, API and Pro.) ### Why this is striking - The same failure as case 006, from another vendor. When not told otherwise, a model names itself after the newest model in its **training data**, which is its own predecessor. - The user read it as the vendor **cheating**. Cutoff blindness about the model itself becomes a trust problem for the product. ### Correction None documented in the issue. ### Lesson Don't treat a model's self-report as evidence of which model you're using. Check the API response's `model` field instead. ### Sources - openai/codex #4488: ## Related cases - GPT-5.5 said its cutoff was June 2024: https://postcutoff.com/why/018-gpt-5-5-reports-june-2024-cutoff/ - Newer Claude models called themselves an older model and reported an old cutoff: https://postcutoff.com/why/006-claude-sonnet-4-says-it-is-claude-3-5-sonnet/ - ChatGPT Voice answered that Kirk “has not been assassinated” and was still alive: https://postcutoff.com/why/017-chatgpt-voice-says-kirk-not-assassinated/