GPT-Live Failed Its First Viral Test in 12 Seconds

OpenAI spent months building GPT-Live to make conversations feel natural. Full duplex, real-time translation, handles interruptions like a person. The internet spent about 12 seconds finding something it can’t do.

On July 8, OpenAI released GPT-Live-1 and GPT-Live-1 mini, replacing Advanced Voice Mode in ChatGPT. The new architecture processes listening and speaking in one model instead of a pipeline, which should make conversations smoother. And by most accounts it does. Simon Willison called it very impressive after a multi-week preview, and OpenAI’s own evaluations show strong preference over the old system across turn-taking, interruption handling, and conversational flow.

Then Husk got hold of it.

How do you break a full-duplex voice model with a spelling question?

Husk is a TikToker whose whole bit is asking ChatGPT questions it should answer easily and watching it fail. He has a track record. He’s gotten previous models to spell “December” with an X and “eighty” with an A. His GPT-Live test was deceptively simple: he asked the model how many letter E’s appear in the word “seventeen”.

GPT-Live answered without hesitation: two. One in “seven” and one in “teen”. It said it with confidence, the kind of confidence that makes you second-guess yourself for a second. When Husk pushed back, the model doubled down. Then it spelled the word correctly, letter by letter, revealing all four E’s, and still insisted there were only two. An OpenAI Codex engineer reposted the video with a single-word caption that summed it up: “FUCK.”

Husk isn’t the only one. Other users ran the classic “strawberry test” (three R’s, not two) with similar results. And GPT-Live’s new full-duplex capability, the feature OpenAI markets most heavily, has its own viral annoyance. The model says “mhm” and “yeah” to show it’s listening, except it can’t stop. Users are posting clips of it interjecting constantly, turning what was meant to feel like a natural conversation into an AI that cannot stop agreeing with you.

Why can’t it count letters in a word?

The short answer is that GPT-Live doesn’t read words the way you do. It uses tokens, chunks of text that don’t map neatly to individual letters. Asking it how many E’s are in “seventeen” is like asking a speed reader for the 15th word on the page they just scanned. The information is there but the model isn’t built to retrieve it that way. It knows the word, it can spell it, it can define it, but counting the letters inside it requires a kind of character-level lookup that transformer architectures don’t do natively.

The full-duplex part is a different problem. The model is trained to produce backchannel cues (mhm, yeah, got it) because human conversations are full of them. But a model that never learned when to stop backchanneling is like a dinner guest who says “wow, really?” after every sentence. It’s technically listening. It’s just exhausting to talk to.

None of this means GPT-Live is bad. The underlying architecture is genuinely the right direction for voice AI, and the translation demos are impressive within their language sweet spot. But the gap between a model’s lab performance and its first day on the internet is always bigger than anyone expects. OpenAI shipped a voice model that can handle real-time interruption, live translation, and background noise filtering. The internet decided to test it on fourth-grade spelling. And it lost.

FAQ

Is GPT-Live available through the OpenAI API?

Not yet. GPT-Live is currently a ChatGPT-only feature. OpenAI has opened a notification form for API availability but hasn’t published pricing or a timeline. The closest API option is the Realtime API with gpt-realtime-2.1.

What is full duplex in voice AI?

Full duplex means a model can listen and generate speech at the same time, like a real conversation. Older systems were half-duplex (walkie-talkie style), where one person talks then stops before the other starts. GPT-Live is OpenAI’s first full-duplex consumer voice model.

Will GPT-Live improve on spelling tests?

The spelling test failures are a tokenization problem, not a reasoning problem. OpenAI could add a character-level verification step, but that’s a UX fix for what is fundamentally a mismatch between how the model processes text and what users expect.

How much does GPT-Live cost?

GPT-Live-1 mini is the default voice mode for free-tier ChatGPT. The full GPT-Live-1 model requires a paid plan starting at $8/month (Go), $20/month (Plus), or $100/month (Pro) for unlimited use.