Retell surpasses 100 million monthly calls using OpenAI models
Retell AI, a voice AI platform for contact centre operations, handled more than 100 million monthly calls by September 2026, rising from 30 million in early 2026. The company says its platform serves more than 4,000 businesses, while OpenAI models generate responses in 80% of calls and customers achieve resolution rates above 80%.
Retell has consumed more than 1 trillion tokens across OpenAI models. It uses the models to power agents that answer questions, take actions through business-system integrations and transfer callers to human staff when judgment is required.
Benchmarking models for live conversations
Enterprise voice agents must follow company rules, remain grounded in supplied information and use tools correctly as a conversation changes. Retell built its vCX benchmark around difficult, multi-turn scenarios drawn from production use cases to compare dozens of models on instruction following, grounding, tool use and conversational quality.
In Retell’s October 13, 2026 evaluation, GPT-6 Astra ranked first for instruction following with 89.7%. Evaluations conducted before that model’s launch put GPT-5.5 first for non-hallucination rate at 88.0% and tool-call correctness at 88.3% with reasoning enabled, ahead of Gemini-3.1-Pro and Claude-4.6 on those measures. The expanding OpenAI model portfolio, reflected in OpenAI DevDay 2026 platform announcements across ChatGPT and Codex, gives developers a broader set of products to assess for production workflows.
Retell makes new OpenAI models available when they launch and reruns its benchmark as models improve. Customers can also use OpenAI Realtime models through the Realtime API and OpenAI text-to-speech as a voice engine.
Resolution, handoffs and operational scale
During a call, Retell’s harness runs audio models, tracks the dialogue and manages turn-taking and interruptions. GPT-6.1 Sol, GPT-6 Astra and GPT-5.5 generate responses and request tool calls using the customer’s instructions and call context. Retell executes those calls through integrations and can write results back to customer records.
When a case needs a person, the agent briefs the representative, answers follow-up questions and confirms readiness before connecting the caller. After the call, Retell uses GPT-5.6 Terra to analyse the transcript and extract outcomes and other data.
Anker uses Retell agents for Tier-1 post-sales support in the United States and United Kingdom, including order tracking, warranty checks and troubleshooting. Anker reported an 80.4% case resolution rate and an NPS of 63. Pine Park Health uses agents for appointment reminders and proactive scheduling; it reported medical-assistant time equivalent to 2.1 full-time roles recovered for patient care and a 55.7% booking rate when the scheduling agent reached a patient.
Latency and engineering workflow
Retell says daytime traffic can exceed its overnight low by more than 50 times. It uses dedicated OpenAI servers shared across customers, and its October benchmark recorded median first-sentence latency of 883 milliseconds for GPT-6 Astra and GPT-6 Sol. GPT-5.6 Luna recorded the highest result in that test at 497 milliseconds.
Regional API endpoints let Retell direct traffic to specific geographies to reduce network travel time and meet data-residency requirements. The company also operates its own telephony stack with a second stack ready to take over, plus backups for voice and model services.
Engineers use Codex to investigate unfamiliar code, build features, debug by cross-referencing logs and database records, and help customers draft model instructions. Retell also uses GPT-6 Astra’s computer-use capability to test front-end pull requests. The company reports an annual revenue run rate of $1.6 million per employee; for businesses, the practical implication is to evaluate voice AI against real call quality, response time, integration accuracy and reliable human handoff before expanding deployment.