You hand your phone to a creator with a reputation for humiliation. I watched GPT-Live-1 stumble on a one-word counting test. For a beat the machine sounded convincingly human—and wrong.
He asked the model how many E’s are in “seventeen” — and it answered “two”
Husk, a TikToker who has become the internet’s single-player red team, whispered a child’s puzzle into OpenAI’s new full-duplex voice model. The clip is three seconds of pure diagnostic cruelty: clear prompt, confident voice, wrong answer. I sat there waiting for the recovery that never came.
Had to give the new voice model the classic test pic.twitter.com/fQYHbBRNuL
— Husk (@huskirl) July 8, 2026
OpenAI framed GPT-Live-1 as a step toward smoother back-and-forths—real-time listening and talking so the assistant can respond mid-sentence. Instead, Husk’s clip exposed a brittle edge in the public demo: fluency without reliable truth. The model’s confident “two” landed like a magician fumbling a coin trick.
People are reusing classic benchmarks and the model keeps failing them
Anyone who has poked at voice assistants keeps a small list of humiliations: timers that won’t set, letter counts that go south, simple paraphrases that melt. Those checks are cheap, replicable, and devastating when they fail on camera.
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
Himelstech gave the model the “strawberry” test, and it tripped again. Hundreds of engineers at OpenAI—Codex folks included—have probably replayed these clips in group chats. Jason Liu, a Developer Experience Engineer tied to OpenAI Codex, reposted Husk’s video and wrote one raw word: FUCK. Even Sam Altman has seen the content surface in his feed; these clips reach the top of the company’s hierarchy faster than any internal bug report.
Full-duplex added a new annoyance: chatter while you speak
Turning on mid-speech confirmations was supposed to make conversations feel alive. Instead, users report an interrupting assistant that keeps buzzing “mhm” and “yeah” even when you don’t want it to.
That extra layer of signal-production makes the system feel attentive, until it feels intrusive—like a hairline crack in a dam, small at first and then impossible to ignore. On X, people joke the model has become an interrupting machine, a small but persistent failure mode that wears on patience far quicker than a single wrong answer.
Why did GPT-Live-1 fail simple counting tasks?
Short answer: the model optimizes for conversational flow and probabilistic text, not the symbolic certainty a human applies to letter-count problems. You and I expect an instant, deterministic answer for seventeen; the model is juggling phonetics, tokenization quirks, and safety layers that can nudge its output away from crisp arithmetic or character-level counts.
Can these problems be fixed without pulling the product?
Yes, but fixes are messy. Engineers at OpenAI will likely patch token-handling, tune the voice stack, and add targeted tests that catch these micro-fails. They have the telemetry—ChatGPT and the GPT APIs already feed usage metrics back to the company—but the tension is product cadence versus public perception. Every viral clip increases pressure from boards, developers, and customers who pay for reliability across Slack, TikTok, and X.
People are using low-effort tests as high-visibility signals
Those three-second micro-fails do more reputational damage than a long technical postmortem. A creator posts a short clip, it reaches millions, and suddenly the narrative about “progress” flips into a narrative about fragility.
As someone who watches models ship and then meets the public fallout, I notice a predictable arc: demos, viral stress tests, PR attempts to downplay, and then a sprint of engineering triage. The voice model’s new skills—live translation and mid-speech responses—are real advances and will matter for accessibility and agents. But when basic tasks break on camera, trust frays faster than features accumulate.
OpenAI has options: throttle releases, add guardrails, or lean into transparency and show test artifacts. You can see their trade-offs in the reactions from the developer community and the blunter reactions from creators like Husk. The question isn’t whether the tech is impressive; it is whether it meets the baseline expectations of everyday interaction.
If you were building a product on GPT-Live-1—on ChatGPT voice integrations in apps or on a TikTok bot—you’d have to ask yourself which risk you can tolerate: a smooth-sounding assistant that occasionally lies, or a quieter tool that answers correctly. Which would you choose?