Voice Mode Was Weak

ChatGPT's Advanced Voice Mode may feel on the surface like the smartest AI around, but it actually runs on an older, weaker model from the GPT-4o era. When asked about its knowledge cutoff, it answers April 2024, revealing that it's a much earlier model than users would expect.

This connects to a point raised in a tweet by Andrej Karpathy about how perceived performance can vary greatly depending on the access path to a model. He noted that while OpenAI's free voice mode can stumble even on trivial questions, the paid Codex model shows far stronger capability, restructuring codebases or finding security vulnerabilities over the course of an hour.

There are two key differences at play.

• Domains with a verifiable reward function are easier to train and improve. For example, coding tasks have clear signals like whether unit tests pass.
• More resources get concentrated in areas with high B2B value. That's why code and security advance faster than consumer experiences like voice.

The upshot is that even within the same OpenAI product lineup, the perceived performance gap can be huge depending on the touchpoint and domain the user encounters.