Retell voice AI platform hits 100 million monthly calls using OpenAI models
Article Summary
Retell AI now processes over 100 million voice agent calls monthly for more than 4,000 businesses, with OpenAI models powering 80% of these interactions. The platform utilizes models such as GPT-6 Astra, GPT-6.1 Sol, and GPT-5.5 to manage complex workflows, achieving customer resolution rates exceeding 80%. According to Retell's October 13, 2026, vCX benchmark, GPT-6 Astra ranked first in instruction following with an 89.7% score, while earlier evaluations showed GPT-5.5 outperforming competitors in non-hallucination and tool-call correctness.
AI Analysis
## Core Assessment
Retell AI has scaled to 100 million monthly calls by effectively abstracting the complexity of live voice interactions into a modular API, relying on OpenAI’s suite of models for 80% of its logic. The reported 80%+ resolution rate is a strong metric for automated support, yet it relies heavily on the platform's ability to manage state and tool-calling accuracy rather than just raw model intelligence. The trade-off here is operational: Retell offloads the model reliability problem to OpenAI while internalizing the telephony stack and latency management—evidenced by their median latency reaching 883 milliseconds for newer models. While impressive, the platform’s dependence on a single primary vendor for the bulk of its inference means their "redundancy" strategy must compensate for API-level outages or model regressions, regardless of how high a model ranks on a custom vCX benchmark.
## Impact and Advice
* **Contact Center Managers:** Do not conflate "resolution rate" with "customer satisfaction." While 80% resolution is a high bar, review the remaining 20% of escalated cases to determine if failures are due to model hallucinations or lack of integration depth.
* **Infrastructure Architects:** Evaluate if your data residency requirements align with the regional API endpoints Retell uses. Ensure your failover strategy accounts for model-provider downtime, not just telephony stack redundancy.
* **Developers:** Use Retell’s provided vCX benchmarks as a starting point, but validate them against your specific production edge cases; a high instruction-following score in a benchmark does not guarantee the same performance in your unique, noisy call environment.
## Contrarian View
Retell’s reliance on "real-time" AI models introduces a persistent latency floor that can frustrate users during high-traffic spikes. Despite claiming to handle 50x demand shifts, the 883ms latency for the top-performing models like GPT-6 Astra is nearing the threshold where human-like conversational fluidity begins to degrade. Reliance on a single vendor’s ecosystem, even with model variety, limits the ability to optimize for specialized, lower-latency local models if API performance fluctuates.
## Bottom Line
Retell has successfully turned voice-AI from a lab experiment into a high-volume utility, but the real engineering heavy lifting remains in managing the brittle interface between LLM reasoning and the messy reality of telephony.
Retell AI has scaled to 100 million monthly calls by effectively abstracting the complexity of live voice interactions into a modular API, relying on OpenAI’s suite of models for 80% of its logic. The reported 80%+ resolution rate is a strong metric for automated support, yet it relies heavily on the platform's ability to manage state and tool-calling accuracy rather than just raw model intelligence. The trade-off here is operational: Retell offloads the model reliability problem to OpenAI while internalizing the telephony stack and latency management—evidenced by their median latency reaching 883 milliseconds for newer models. While impressive, the platform’s dependence on a single primary vendor for the bulk of its inference means their "redundancy" strategy must compensate for API-level outages or model regressions, regardless of how high a model ranks on a custom vCX benchmark.
## Impact and Advice
* **Contact Center Managers:** Do not conflate "resolution rate" with "customer satisfaction." While 80% resolution is a high bar, review the remaining 20% of escalated cases to determine if failures are due to model hallucinations or lack of integration depth.
* **Infrastructure Architects:** Evaluate if your data residency requirements align with the regional API endpoints Retell uses. Ensure your failover strategy accounts for model-provider downtime, not just telephony stack redundancy.
* **Developers:** Use Retell’s provided vCX benchmarks as a starting point, but validate them against your specific production edge cases; a high instruction-following score in a benchmark does not guarantee the same performance in your unique, noisy call environment.
## Contrarian View
Retell’s reliance on "real-time" AI models introduces a persistent latency floor that can frustrate users during high-traffic spikes. Despite claiming to handle 50x demand shifts, the 883ms latency for the top-performing models like GPT-6 Astra is nearing the threshold where human-like conversational fluidity begins to degrade. Reliance on a single vendor’s ecosystem, even with model variety, limits the ability to optimize for specialized, lower-latency local models if API performance fluctuates.
## Bottom Line
Retell has successfully turned voice-AI from a lab experiment into a high-volume utility, but the real engineering heavy lifting remains in managing the brittle interface between LLM reasoning and the messy reality of telephony.