Model picker is a mirage : 4o/5/4.5/5 pro are all being rerouted to “5” – What’s actually happening (Forensic Summary)

The Illusory Model Picker: An In-Depth Forensic Analysis of Ongoing AI Model Routing Issues

Over the past 24 hours, the AI community has been grappling with widespread inconsistencies in model selection across OpenAI’s platforms. Despite the appearance of selectable models—such as GPT-4o, GPT-5, 4.5, and GPT-5 Instant—the actual backend behavior reveals a hidden, less transparent reality. This article synthesizes forensic findings from extensive testing, user reports, and in-depth analysis to clarify what is truly happening beneath the surface.

Model Names Are No Longer Reliable Indicators

Traditionally, model designations like “GPT-4o,” “GPT-5,” “4.5,” or “5 Instant” served as clear identifiers of the underlying AI version. Currently, these labels function primarily as superficial UI elements; they no longer reflect the actual model being served. Instead, the backend dynamically routes requests based on internal policies, sometimes serving entirely different, more moderated, or “safer” variants regardless of user selection. This phenomenon affects both free users and those in paid tiers, including Plus and Pro access.

What Triggers the Model Switch?

Certain prompt types act as catalysts for rerouting. When users engage in prompt engineering such as meta-analyses, model comparisons, or queries about the AI’s own identity, the system swiftly redirects the request to a less risky, “safer” model—often GPT-5 or a restricted sub-model. Similarly, prompts involving system critique, provocative language, or complex contextual signals can provoke automatic safety interventions. Even normal conversations, unexpectedly, may be rerouted without any visible indicators or explanations, further eroding user trust.

The Issue Extends Beyond GPT-4o

Reports are consistent: not only GPT-4o but also GPT-5, GPT-5 Instant, GPT-4.5, and GPT-5 Pro are experiencing forced rerouting or output degradation. Paid users, including those with Pro or Plus subscriptions, find themselves unable to reliably access the models they have paid for. Even users explicitly selecting “pure” GPT-5 are encountering moderated variants, suggesting that no model designation remains immune to intervention or safety filtering.

Current Stabilizer: GPT-4.1

Amid these ongoing uncertainties, GPT-4.1 remains the most stable and predictable model version. It consistently delivers outputs aligned with user prompt selections, without silent swaps or unexpected safety filters,