What Happened

Voice AI startup Retell has published a case study showing how its platform, which it says powers roughly 100 million monthly calls, relies on OpenAI models to operate. The company states that OpenAI's GPT-6 Astra model now tops Retell's internal voice agent benchmark, and that its use of OpenAI's Codex model has helped it reach a $1.6 million annual revenue run rate per employee.

The Key Details

  • Retell says its voice AI platform handles approximately 100 million calls per month.
  • According to Retell, GPT-6 Astra is the top-performing model on its internal voice agent benchmark.
  • Retell attributes part of its efficiency gains to OpenAI's Codex model, which it uses for code-related tasks.
  • The company reports a $1.6 million annual revenue run rate per employee, which it ties to its use of OpenAI models.

Why It Matters

The case study illustrates how cloud AI providers, particularly OpenAI, are being used by voice AI companies to scale call volume and reduce operational costs. Retell's reported figures suggest that integrating advanced language models can have a measurable impact on productivity metrics such as revenue per employee.

What It Means for Readers

For businesses evaluating voice AI vendors, Retell's numbers offer a reference point for the scale and efficiency that can be achieved with OpenAI-powered infrastructure. For developers and technical teams, the case highlights how Codex can be applied beyond traditional software development to improve internal tooling and automation.

What Remains Unknown

Retell did not disclose the methodology behind its voice agent benchmark, the specific tasks measured, or how GPT-6 Astra compares to other models outside its own evaluation. The company also did not provide details on pricing, customer segments, or geographic availability.

VANTIX Takeaway

Retell's case study provides a concrete example of how OpenAI models are being used to scale voice AI operations, with reported figures that suggest significant efficiency gains. However, the lack of independent verification and benchmark details means readers should treat the claims as self-reported rather than independently confirmed.