Retell AI is a voice AI platform for contact centers that now serves more than 4,000 businesses. Its monthly call volume rose from about 30 million in early 2026 to over 100 million by September 2026. To maintain call quality and reliability as usage grows, Retell relies extensively on OpenAI models across production, testing and development.
Key metrics and outcomes
- OpenAI models generate responses for 80% of calls on Retell’s platform.
- Customer resolution rates are reported above 80%.
- Retell has consumed more than 1 trillion tokens across OpenAI models, indicating broad adoption.
Which models are used and for what
Retell uses multiple OpenAI models for specific roles:
- GPT‑6 Astra: used for front‑end testing and for the company’s vCX (voice customer experience) benchmark; in Retell’s October 13, 2026 benchmark it ranked first for instruction following with 89.7%.
- GPT‑6.1 Sol and GPT‑5.6 Terra: used for live conversation generation and post‑call transcript analysis respectively.
- GPT‑5.5: in pre‑Astra evaluations it achieved the highest non‑hallucination rate (88.0%) and tool‑call correctness (88.3%) with reasoning enabled, outperforming Gemini‑3.1‑Pro and Claude‑4.6 on those measures.
- Codex: used throughout development for code research, feature construction, debugging, and for creating the instructions that guide models in customer applications.
The vCX benchmark
Retell built the vCX benchmark to evaluate how models handle hard, multi‑turn conversations drawn from real production scenarios. The benchmark examines:
- instruction following,
- grounding in provided information,
- correct tool calling,
- natural conversational language.
Retell reruns the benchmark as models improve and makes new OpenAI models available to customers on their launch day.
Business and operational impacts
- Customer examples: Anker uses Retell agents for Tier‑1 post‑sales support in the US and UK (order tracking, warranty checks, troubleshooting). Anker reported an 80.4% case resolution rate and a Net Promoter Score of 63, and said agent tone, pacing and warmth met expectations for human agents.
- Pine Park Health: Retell handles appointment reminders and proactive scheduling, writes outcomes to patient charts, and recovered medical‑assistant time equivalent to 2.1 full‑time roles; the scheduling agent achieved a 55.7% booking rate when reaching a patient.
Performance, latency and reliability
- Peak daytime traffic can exceed overnight low traffic by more than 50x. To keep calls responsive at peak demand, Retell uses dedicated servers with OpenAI shared across customers.
- In the October 13, 2026 benchmark, GPT‑6 Astra and GPT‑6 Sol recorded a median first‑sentence latency of 883 milliseconds; GPT‑5.6 Luna was fastest at 497 milliseconds.
- Retell leverages OpenAI’s regional API endpoints to route traffic to specific geographies, reducing network travel time and helping meet data residency requirements in regulated industries.
- Redundancy is built into each layer: Retell runs its own telephony stack with a standby stack, and maintains backups for voice and model services to keep calls running when a service fails.
How engineers use Codex and GPT models
As enterprise customers request more custom workflows and integrations, Retell engineers often work in unfamiliar code areas. Their workflow includes:
- Providing Codex with a high‑level requirement.
- Having Codex survey the codebase, compare it with real data, and explain required changes.
- Using Codex during implementation to catch overlooked details and during debugging to cross‑reference logs and database records with code.
Retell says Codex often does this research faster and in greater detail than a developer could on their own. Engineers also use Codex to speed up writing the model instructions that guide customer applications.
For front‑end testing, Retell relies on GPT‑6 Astra’s computer‑use capability: when an engineer opens a front‑end pull request, GPT‑6 Astra opens the interface, clicks through flows, attempts to break them and reports issues. Retell reports earlier models could not perform this reliably.
Productivity and future plans
- Retell reports an annual revenue run rate per employee of $1.6 million.
- Planned improvements include expanded quality assurance and simulation, broader language support, and knowledge‑base auto‑learning.
Conclusion
By integrating OpenAI models across production and engineering workflows, Retell has scaled its voice AI platform to serve millions of callers while maintaining high resolution rates and operational safeguards. The company aims to enable its customers to own and operate tailored voice agents internally—"No one knows a company’s customers better than the company itself. We give them the power to act on that expertise and improve AI agents directly. The goal is for them to build an in‑house team of internally deployed engineers who own and run it all," a company leader said.
