Yesterday, OpenAI released its latest voice model, GPT-Live-1. The company called it “a new generation of voice models for natural human-AI interaction.” The full-duplex model allows ChatGPT to talk and listen at the same time, giving the back-and-forth a more conversational flow. It’s designed to be smarter, faster, and more natural.
Sounds great! Let’s see how it handles the internet’s one-man voice model red team, TikToker @huskistaken (aka Husk), when he tests out its capabilities by giving the model something simple:
Had to give the new voice model the classic test pic.twitter.com/fQYHbBRNuL
— Husk (@huskirl) July 8, 2026
Ooh, so close. Husk gave ChatGPT the straightforward request to tell him how many times the letter E appears in the number “seventeen,” and the voice model managed to get halfway there by answering “two.” It then offered a very awkward sign-off when Husk ended the conversation. Maybe that’s a query that GPT-Live-2 will be able to handle.
To be fair, Husk isn’t really testing any of the new features that OpenAI highlighted with the release of its new model, which, by most accounts, seems like it’s more capable when it comes to things like live translation than previous iterations. But he is giving the model a very simple benchmarking test that, despite being OpenAI’s flagship voice model, it still failed spectacularly.
It seems OpenAI knows just how bad this looks, too. Jason Liu, a Developer Experience Engineer on OpenAI Codex, reposted the video with a very succinct, “FUCK.”
Husk’s made a habit of terrorizing OpenAI over its voice model for some time now, making viral videos out of how poorly it handles certain prompts. One of his videos—in which he tasks ChatGPT with setting a timer and watching it fail—made it all the way to CEO Sam Altman, who tried to laugh it off in a very “I’m not mad, please don’t put in the newspaper that I got mad” kinda way.
Husk is not alone in giving GPT-Live-1 the old stress test. Another user on X gave the model the “Strawberry” test, asking it to count the number of times the letter R appears in the word “strawberry.” It’s a classic, and one that most AI models can answer correctly by now (whether that’s because they really know there are 3 Rs or because they’ve been trained on how to respond to prevent embarrassment is another question), but GPT-Live-1 gets tripped up by it.
OpenAI’s brand new voice model vs counting the r’s in strawberry. I really tried to help it along. pic.twitter.com/Pm0RfLyxIA
— Himelstech (@himelstech) July 9, 2026
Another user noted a new annoyance that has arisen from the model’s full-duplex functionality: While the model now allows ChatGPT to say things like “mhm” and “yeah” while you talk to confirm it is listening, it apparently is incapable of simply shutting up and not doing that. It seems ChatGPT has built an interrupting machine. Just what we’ve all been waiting for.
READ MORE:
The Future Is Always Listening: OpenAI Says Its New Voice Assistant Is ‘One Step Closer to a Truly Accessible AGI’
OpenAI Is Tired of Seeing All Those Videos of People Clowning on Its Voice Mode
Read the full article here
