GPT-Live-1 tested on real phone calls: natural speech but serious instruction-following flaws
kolchinski · reddit · 2026-09-14
ThunderPhone (an AI phone-agent company) ran OpenAI's new full-duplex speech-to-speech model GPT-Live-1 on a real phone number with a 13k-token insurance qualification script, across a dozen real calls and 25 simulated ones.
The good: the most natural phone conversationalist they've deployed—one continuous audio stream, 1.3s to first audio, zero VAD or turn-taking logic needed; interruptions, backchannels, and mid-sentence corrections just work.
B2B problems:
- Literalism: prompts ignored—caller says "I own it now," model still reads the own/rent/parents menu; held across every prompt variant.
- "Yes" handled as "no" multiple times, re-asking or moving on as if declined.
- Thinks out loud to callers under pressure ("Hmm. Handling this one carefully...").
- Alphanumeric errors: added a letter when reading back a claim number (audio wrong, transcript right), swapped Q for X in Russian (both wrong).
- Non-English is fluent but heavily American-accented; one Luganda call was answered in Swahili.
The founder promises full transcripts and audio clips in the comments.
Related event: GPT-Live-1 Real-World Test: Most Natural Voice, Shaky Instruction-Following(2 posts)→
More from Models
- Leak: Google Reportedly Has Another 'Hella Chonky' Model After Carbon — iruletheworldmo · 2026-10-30
- Users report across-the-board GPT quality drop: weaker writing and broken instruction following — Kyrannio · 2026-10-30
- Leak claims Gemini Argon hits Pro trial from November 2nd (unverified) — Emotional-Company847 · 2026-10-30
- OpenAI retiring waves of legacy models: GPT-5 family dies Dec 11, GPT-4.1 by Jan 2027 — ThePeterMick · 2026-10-30
- DeepMind CEO Demis Hassabis tells Musk: testing with trusted partners, coming soon — demishassabis · 2026-10-30
- Adobe stands firm on not training on customer data, using licensed stock instead — omooretweets · 2026-10-30