Stealth Model Swap? GPT-5.5 High Claims Knowledge Cutoff is June 2024

First of all, I have tremendous respect for your deep insight into LLM architectures . It’s clear from your words that you have a profound understanding of how these underlying token factories operate. I’d like to use my limited knowledge to piece together your explanation and see if I’m understanding it correctly:

1. When the model tells me its training cutoff is June 2024, you believe this is a hallucination.

2. In a completely isolated session where I banned internet browsing, it stated the latest OpenAI model it knows of is GPT-4o (released in May 2024). You believe this is also a hallucination.

So, the model hallucinated the exact same historical timeline and identical year/month constraints across two completely independent tests. In your view, this is just a highly precise coincidence, correct?

I completely agree with your point that LLMs lack a native self-awareness of the hardware or specific clusters they run on. However, as developers, we both know that Codex and most production-level Agent workflows inject metadata via system prompts into the session context to define the environment.

When I poked further into the session metadata, Codex explicitly spat out that its current environment is “Codex based on GPT-5.0” —whose actual training data timeline happens to align closely with mid-2024.

Are we still calling this a triple-hallucination-coincidence?

To me, this no longer looks like a token factory spinning wild fantasies. It looks like a classic backend misconfiguration where the routing layer is pulling from a legacy GPT-5.0 container or an old Codex base, even though the frontend UI proudly displays “GPT-5.5 High.”