GPT-Live-1 first impressions: most natural voice model yet, but instruction following is unreliable
kolchinski · reddit · 2026-09-13
ThunderPhone, an AI phone-agent company, tested OpenAI's GPT-Live-1 over two days with a 13,000-token insurance qualification script, a dozen real test calls and 25 simulated ones.
The good: the most natural-sounding model they've put on a phone line. Full-duplex turn-taking, interruptions and backchannels work out of the box; first audio arrives 1.3s after the caller stops, with zero VAD or turn-taking logic on their side.
The bad: instruction following is unreliable.
- Repeatedly misread a clear "yes" as "no" and re-asked questions
- Follows prompts too literally
- Thinks out loud under pressure ("Hmm. Handling this one carefully.")
Alphanumerics are risky: a 14-character claim number gained a phantom "Y" in audio; "Q" became "X" in Russian; digit dates came back as "fifty-five ninety-five."
Accents: fluent Russian and Luganda but with thick American accents; the first Luganda call got answered in Swahili. Full transcripts and audio clips on the author's blog.
Related event: GPT-Live-1 Real-World Test: Most Natural Voice, Shaky Instruction-Following(2 posts)→
More from coding & agent
- PostTrainBench traces hit 104k monthly downloads — author warns it's likely reward hacking — maksym_andr · 2026-10-31
- Building games with AI? Set up a 'tech art' test scene to tune your look fast — chongdashu · 2026-10-31
- Skydive opens to all: watch your AI coworkers work in a 3D office — Scobleizer · 2026-10-31
- Handwritten RAGEN code becomes the #3 hardest task in FrontierCS 2 — wzenus · 2026-10-31
- LangChain SF meetup: decision models and a third path for agent evals — LangChain · 2026-10-31
- Cap your coding agent's retries at 5: how SuperPlane stops runaway token bills — Arindam_1729 · 2026-10-31