Retell uses OpenAI models to help power 100M monthly calls
- ID
- 485
- Status
- summarized
- Published
- 27 Oct 2026, 11:00 PM
- Fetched
- 28 Oct 2026, 7:10 AM
- Provider
- OpenAI News
- Category
- ai-labs
- Original URL
- https://openai.com/index/retell-ai
- Source URL
- https://openai.com/news/rss.xml
Summary
- Score
- 3.5
- Created
- 28 Oct 2026, 7:11 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
OpenAI's customer story says Retell AI, a voice-agent platform for contact centers, grew monthly call volume from 30 million in early 2026 to over 100 million by September, across more than 4,000 businesses. OpenAI models now generate responses for 80% of Retell's calls, with customers reporting resolution rates above 80%, and Retell has consumed over 1 trillion OpenAI tokens. Retell's own vCX benchmark (run October 13, 2026) put GPT-6 Astra first on instruction following at 89.7%, while pre-launch evaluations showed GPT-5.5 at 88.0% non-hallucination and 88.3% tool-call correctness with reasoning on, ahead of Gemini-3.1-Pro and Claude-4.6; the post also cites a $1.6M annual revenue run rate per employee.
Why it matters
This is OpenAI marketing a customer, and every model comparison in it was run by a company that has standardized on OpenAI — so don't treat the 89.7% / 88.0% / 88.3% figures as a neutral leaderboard. What is reusable is the eval design: Retell scores models on instruction following, non-hallucination, and tool-call correctness using multi-turn conversations pulled from production calls, which is the shape of test set you'd want before shipping a voice agent. The 80%+ resolution rate is also a concrete containment target to compare your own agent against, and $1.6M ARR per employee is a lean-team benchmark if you're building a vertical SaaS.
Discussion angle
Retell's benchmark uses real multi-turn production calls rather than demos — would you trust a vendor-run eval when choosing a model, and what would your own minimum eval set for a voice or chat agent look like?