OpenAI Launches GPT-Live with Full-Duplex Voice, But Counting Errors Emerge

OpenAI has launched GPT-Live, a new generation of voice models for ChatGPT that can listen and speak simultaneously, aiming for more natural conversations. The full-duplex architecture processes audio continuously, allowing interruptions and real-time backchanneling. However, viral tests show the model still struggles with basic counting tasks.

What is GPT-Live

• GPT-Live uses a full-duplex architecture, enabling the AI to listen and talk at the same time Source: La Vanguardia.
• It supports interruptions, active listening cues like "ajá," and can display visual cards Source: Infobae.
• The system delegates complex tasks to GPT-5.5 in the background while maintaining conversation flow Source: Infobae.

Counting Failures

Despite the advanced voice capabilities, GPT-Live-1 incorrectly counted letters in simple words:

• Asked how many 'e's are in "seventeen" (four), it answered two Source: El Confidencial.
• Another test with "strawberry" also resulted in a wrong count Source: El Confidencial.
• The developer Jason Liu reacted with "Joder" (a Spanish expletive) to the error Source: El Confidencial.

What's Next

OpenAI plans to expand GPT-Live to API and enterprise users. The company emphasizes safety measures, including parental controls and fraud prevention Source: Infobae. Despite the letter-counting hiccups, GPT-Live represents a significant step toward more human-like AI interaction.