ChatGPT's voice mode told a user that its base model was GPT-6, and in the same conversation failed to count the letter e in the word "seventeen." An OpenAI employee later said the two halves were connected: Voice actually runs on an older model, GPT-Live-1, so the assistant was wrong about itself as well as about the spelling. The exchange was posted by a user named Husk. It is a trivial failure sitting on top of a disclosure problem that is not trivial at all.
The question was as simple as questions get: how many e's are in "seventeen." There are four. ChatGPT said three. It had just told Husk that its base model, GPT-6, was "generally" considered one of the smartest available. Then it tried to check itself by spelling the word out loud — S-e-v-e-n-t-e-e-n — and concluded, again, that there were three.
When Husk pointed out the error, the bot did not revise. It repeated, sharply, that it had said three and that three was correct. By Husk's count, the assistant had by then pronounced four e's roughly five times while arguing that there were three. Eventually it appeared to give up.
The clip was captioned with a reference to the thermal exhaust port from Star Wars — the small unguarded opening that brings down the whole station.
An OpenAI employee replied to Husk that ChatGPT Voice, contrary to what the bot said about itself, still runs on GPT-Live-1. Voice can hand a request off to a more capable model when the situation calls for it. In this conversation, it did not.
Husk asked the obvious follow-up: if a user explicitly disputes an answer and asks for it to be rechecked, why isn't that the trigger for a handoff? The OpenAI employee acknowledged it was a good question.
It is a good question, and the answer is presumably cost. Routing is where the economics of a consumer assistant live: the cheap model takes the traffic, the expensive model is reserved for what looks hard. A letter-counting question does not look hard. That is precisely why this case is instructive — the router's estimate of difficulty and the model's actual competence came apart, and nothing in the loop was built to notice.
What interests me more than the wrong count is the wrong self-description. A miscount of letters is a known artifact of how these systems read text, and almost everyone reading this has seen one. A model confidently naming a version of itself that it is not running is a different category of error, and it is the one users have no way to catch. The stated defense of GPT-6 here also doubles as an indictment: if the assistant cannot reliably report which model is answering, then every claim a user makes about "what GPT-6 said" is unverifiable, in both directions.
That matters because of who this product is aimed at. OpenAI markets voice AI for, among other things, talking with elderly relatives — users least likely to know that a chatbot is a routing layer over several models of different capability, and least likely to push back five times on a confident wrong answer. Husk kept arguing. Most people accept the first reply and move on, which is the entire commercial appeal of routing cheap.
The fix Husk proposed costs almost nothing: treat user disagreement as a signal to escalate. That it isn't already the rule suggests the system is tuned to defend its answers rather than re-examine them — a preference that is cheap to run and expensive to trust.