- Retell AI’s platform grew from 30 million to over 100 million monthly calls between early 2026 and September 2026, according to an OpenAI customer story, with OpenAI models handling an estimated 80% of calls and customer resolution rates reported above 80%.
- Retell’s modular architecture delivers sub-second latency by orchestrating telephony, ASR, LLMs and TTS layers, with GPT-5.6 Luna recording a median first-sentence latency of 497 milliseconds in October 2026 benchmark data.
- Retell Assure, launched January 1, 2026, monitors 100% of calls and automatically refines models in real time, eliminating the spot-check lag that typically delays issue detection by weeks in high-volume voice deployments.
By September 2026, Retell AI was processing over 100 million calls a month across more than 4,000 businesses, with OpenAI models generating responses for an estimated 80% of those interactions. The scale matters because it pushes voice AI past the proof-of-concept stage: these are production deployments handling Tier-1 support, appointment scheduling and outbound campaigns at volumes that expose every latency and reliability flaw in the stack.
From 30 Million to 100 Million
Retell’s call volume tripled in roughly eight months, according to an OpenAI announcement dated October 27, 2026. Sustaining that growth required consuming over 1 trillion tokens across OpenAI’s model family, a figure that reflects the computational weight of real-time voice at scale.
The business results cited by Retell’s customers give some shape to what that throughput produces. Anker, the electronics brand, deployed Retell agents for Tier-1 post-sales support across the US and UK and reports an 80.4% case resolution rate and a Net Promoter Score of 63. Pine Park Health uses Retell for appointment reminders and proactive scheduling, freeing staff time equivalent to 2.1 full-time roles, which the organisation redirected to patient care.
The Real-Time Voice Stack
Latency is the defining constraint in voice AI. Retell’s architecture orchestrates telephony, ASR, LLMs and TTS engines to hold response times below one second, a threshold where conversational naturalness breaks down if crossed. The proprietary turn-taking model at the centre of that stack predicts conversational flow and manages speaker transitions, reading cues like tone shifts and sentence patterns to avoid the dead air that makes synthetic voices feel mechanical.
The platform supports GPT, Claude and Gemini on the LLM side, alongside TTS providers including ElevenLabs and OpenAI, giving engineering teams the option to trade off cost, quality and speed per use case. Retell’s October 13, 2026 benchmark data puts GPT-5.6 Luna at a median first-sentence latency of 497 milliseconds, while GPT-6 Astra and GPT-6 Sol recorded 883 milliseconds. On function-calling reliability, OpenAI’s GPT-4o previously showed a success rate above 70% in multi-turn interactions on Retell, according to the company, nearly double that of earlier alternatives.
GPT-6 Astra in the Dev Loop
Retell uses GPT-6 Astra not only for live customer interactions but inside its own development pipeline. When an engineer opens a front-end pull request, GPT-6 Astra autonomously navigates the interface, clicks through flows, identifies issues and files bug reports. Retell states earlier models could not perform these tasks reliably, making the capability a practical quality gate rather than a benchmark exercise. The pattern points toward a broader deployment model where frontier LLMs serve as active components in CI/CD pipelines, not just generation tools.
Codex and Engineering Output
Retell’s engineering team runs OpenAI’s Codex as a standard part of the development workflow, using it for research on unfamiliar code, feature builds and debugging. The company reports $1.6 million in annual revenue run rate per employee, a figure it attributes partly to this AI-assisted workflow.
In practice, Retell’s engineers report that Codex surfaces overlooked details during implementation and cross-references logs and database records against the codebase when investigating root causes. It also accelerates customer-facing engineering work, including writing the LLM instructions that govern how a model behaves inside a specific customer’s application. For a team scaling rapidly without proportional headcount growth, that kind of leverage on routine engineering tasks is where OpenAI’s model economics translate directly into margin.
Retell Assure: Automated QA
Traditional QA on voice AI relies on humans spot-checking a fraction of recorded calls, a process that routinely takes weeks to surface and fix model issues. At 100 million monthly calls, that lag is operationally untenable.
Retell Assure, launched January 1, 2026, monitors 100% of calls and automatically refines voice AI models in real time. For enterprise customers in regulated sectors, the platform also carries SOC 2 Type II certification, HIPAA compliance with Business Associate Agreements and GDPR adherence. The compliance layer matters especially in healthcare and financial services, where gaps in AI agent monitoring have drawn increasing scrutiny. Removing the human bottleneck from QA is what makes continuous deployment viable at this call volume.
Developer Control at the Core
Retell’s commercial positioning is explicitly developer-first. Rather than a bundled no-code product, it offers a modular system where engineering teams select their LLM, TTS and ASR providers and orchestrate them through Retell’s core engine. That modularity is what enterprise integrations typically require: function calling that lets agents book appointments, update CRM records or transfer to a human agent mid-call, with all of it wired to existing backend APIs.
The platform also runs batch outbound campaigns for lead qualification and proactive engagement, extending beyond inbound call handling. The combination of low-latency infrastructure, developer-configurable components and automated QA is what Retell is using to differentiate from platforms that bundle the stack and limit customisation. At 100 million monthly calls, the architecture has at least been tested at scale.
