Retell 100 Million Calls a Month, Powered by OpenAI
Key Takeaways
- Retell AI now serves more than 4,000 businesses, with monthly call volume growing from 30 million in early 2026 to more than 100 million by September 2026.
- OpenAI models generate responses for 80% of calls on Retell, and Retell reports customer resolution rates above 80%.
- Retell has surpassed 1 trillion tokens consumed across OpenAI models, according to OpenAI’s customer story published October 27, 2026.
- In Retell’s October 13, 2026 vCX benchmark, GPT-6 Astra ranked first for instruction following at 89.7%, a Retell benchmark result, not an independent finding.
- Latency matters on live calls: GPT-6 Astra and GPT-6 Sol recorded median first-sentence latency of 883 milliseconds, while GPT-5.6 Luna recorded 497 milliseconds.
Voice AI has a brutal test that text chatbots do not face: the caller is waiting, in real time, and any hesitation is audible. Retell AI, a voice platform for contact-centre operations, says it has now passed that test at scale. In an OpenAI customer story published October 27, 2026, Retell reports serving more than 4,000 businesses, with monthly call volume growing from 30 million in early 2026 to more than 100 million by September 2026. OpenAI models now generate responses for 80% of calls on the platform.
Those headline numbers come from the companies involved, and they should be read with that in mind. Retell reports customer resolution rates above 80%, and it has surpassed 1 trillion tokens consumed across OpenAI models. Neither figure has been independently audited in the source reviewed. What makes the story useful is not the round numbers alone, but the detail about how voice agents are built, tested and handed over to humans when they should be.
How Retell builds a voice agent
Retell is a platform for building agents that answer questions, take action and connect a caller with a person when needed. Contact-centre teams define an agent’s business context, goals and rules, and developers can configure workflows and integrations. During a call, Retell’s harness runs the audio models, tracks the conversation and manages when the agent speaks, listens or responds to interruptions.
OpenAI models, including GPT-6.1 Sol, GPT-6 Astra and GPT-5.5, use the customer’s instructions and conversation context to generate responses and request tool calls. Retell executes those calls through integrations with the customer’s systems and can write outcomes back to a customer record. After a call, Retell uses GPT-5.6 Terra to analyse the transcript and extract outcomes and other data. The architecture is therefore a pipeline — audio, model, tools, record — rather than a single model answering in isolation.
Human handover is built into that pipeline. When human judgement is needed, the voice agent briefs the representative with context, answers follow-up questions and confirms the representative is ready before connecting the customer. That design choice matters for trust: a voice agent that refuses to hand over, or hands over without context, will frustrate callers even if it resolves most cases.
What the benchmark does and does not prove
Retell built its vCX benchmark to compare dozens of models using difficult, multi-turn conversations drawn from real production scenarios. It tests how well models follow instructions, stay grounded in the information they are given, call the right tools and use natural conversational language. In Retell’s October 13, 2026 benchmark, GPT-6 Astra ranked first for instruction following at 89.7%.
That result is a Retell benchmark result, not an independent evaluation. Retell also shared earlier evaluations, conducted before the launch of GPT-6 Astra, in which GPT-5.5 ranked highest among models on both non-hallucination rate (88.0%) and tool-call correctness (88.3%) with reasoning enabled, outperforming Gemini-3.1-Pro and Claude-4.6 on both measures. Again, those are Retell-reported results, using Retell’s own test set and methodology, and they should not be quoted as neutral industry rankings.
Latency results from the same benchmark are more concrete, because they measure something observable. GPT-6 Astra and GPT-6 Sol each recorded median first-sentence latency of 883 milliseconds, with GPT-5.6 Luna ranking highest at 497 milliseconds. On a live call, the difference between half a second and nearly a second before the first sentence is noticeable, and it illustrates a trade-off: the model that follows instructions best is not necessarily the fastest to start speaking.
Customer examples, with the right caveats
At electronics company Anker, Retell agents handle 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 a Net Promoter Score of 63, and its team said the agents’ tone, pacing and warmth met the standard they expected of a human agent. Those are customer-reported figures, shared via OpenAI’s story, not independently verified outcomes.
At Pine Park Health, Retell agents handle appointment reminders and proactive scheduling and record outcomes in patients’ medical charts. The team recovered medical-assistant time equivalent to 2.1 full-time roles and redirected it to patient care, and its scheduling agent achieved a 55.7% booking rate when it reached a patient. The healthcare example also shows why data residency and routing matter: OpenAI’s regional API endpoints allow developers to route traffic to specific regions, which helps Retell reduce network travel time and meet data residency requirements in healthcare and financial services.
Retell also uses Codex throughout development, from researching unfamiliar code to building features and debugging, and GPT-6 Astra to test front-end interfaces. It reports $1.6 million in annual revenue run rate per employee, a company-reported efficiency figure that reflects a small team relative to call volume, not a profitability statement.
Limitations and unanswered questions
The central limitation is provenance. Call volume, resolution rates, token consumption, benchmark rankings and customer outcomes in this article are reported by Retell or OpenAI, not independently audited. The vCX benchmark is Retell’s own test, and its rankings may not transfer to other companies’ call types, languages or compliance requirements. Daytime traffic can reach more than 50 times the overnight low, according to the story, which raises questions about cost and reliability at peak that the headline volume figure does not answer. Businesses considering voice agents should test them on their own call types, with their own escalation rules, before treating any vendor benchmark as a purchase decision.
For readers, the speech angle connects to NVIDIA’s Nemotron Transcribe open speech model, which shows how speech technology itself is becoming open and multilingual, while Socure’s agent trust scores address the parallel question of whether an AI agent can be trusted with consequential actions. Voice agents raise both questions at once, in real time.
