Retell has scaled its voice AI platform rapidly: by September 2026 the company was handling more than 100 million monthly calls, using OpenAI models to generate roughly 80% of live responses while relying on Codex to accelerate engineering and development. The company, which now supports more than 4,000 businesses, expanded from about 30 million monthly calls earlier in 2026 to its current volume by concentrating on model performance, latency, and operational redundancy.
To evaluate model behavior in production-like conditions, Retell developed a proprietary vCX benchmark that uses multi-turn conversations drawn from real scenarios. The benchmark measures instruction following, grounding, tool-calling accuracy and conversational fluency. In Retell’s October 13, 2026 run, GPT-6 Astra led instruction following with an 89.7% score. Earlier comparisons completed before GPT-6 Astra’s release showed GPT-5.5 posting the highest non-hallucination rate (88.0%) and the best tool-call correctness (88.3%) when reasoning was enabled. Retell re-runs this benchmark as models are updated and makes new OpenAI models available to customers on launch day.
In live customer calls, Retell’s harness coordinates audio models, maintains conversation state and orchestrates agent actions — including speaking, listening and yielding to interruptions. Models such as GPT-6.1 Sol, GPT-6 Astra and GPT-5.5 use customer-defined instructions and the conversation context to generate responses and request tool calls. Retell then executes those tool calls through integrations with customers’ systems and can write results back to customer records. When human oversight is needed, the platform briefs a representative with context and confirms readiness before connecting the caller. After calls, Retell uses GPT-5.6 Terra to analyze transcripts and extract outcomes and structured data.
Customers report measurable benefits from this approach. Electronics maker Anker deploys Retell agents for Tier-1 post-sales support in the US and UK, covering order tracking, warranty checks and troubleshooting. Anker reported an 80.4% case resolution rate and a Net Promoter Score of 63, and said agent tone and pacing met expectations for human-like interaction. In healthcare, Pine Park Health uses Retell for appointment reminders and proactive scheduling, recording outcomes in medical charts; the provider said it recovered the equivalent of 2.1 full-time medical-assistant roles and achieved a 55.7% booking rate on scheduling calls.
Maintaining responsiveness at large scale has required attention to latency and routing. Retell uses dedicated OpenAI servers shared across customers and leverages OpenAI’s regional API endpoints to reduce network travel time and to meet data residency needs in regulated industries. In the company’s October 13, 2026 latency measurements, GPT-6 Astra and GPT-6 Sol recorded median first-sentence latencies of 883 milliseconds, while GPT-5.6 Luna showed a median first-sentence latency of 497 milliseconds. Retell also builds redundancy across its stack by running its own telephony infrastructure with a standby stack and maintaining backups for voice and model services to preserve call continuity if any service fails.
Beyond runtime use, Retell treats Codex as an engineering teammate. Engineers use Codex to survey unfamiliar code, research changes, produce implementation suggestions and debug. Retell reports that Codex can inspect code and related data, explain necessary changes and identify edge details that improve quality. For web interface testing, the company leverages GPT-6 Astra’s computer-use capabilities to open interfaces, click through flows, try to break them and report issues — capabilities Retell says earlier models could not perform reliably. Codex also helps speed tasks such as authoring the instructions that guide model behavior in customer applications. The company reports an annual revenue run rate of $1.6 million per employee.
Retell’s roadmap focuses on giving customers control to test and improve agents: expanding quality assurance and simulation, adding support for more languages and developing knowledge-base auto-learning. The company continues to offer Realtime and text-to-speech models through OpenAI’s APIs and prioritizes model selection based on real-call performance rather than demos, according to co-founder and CEO Bing Wu.
As enterprises adopt voice AI at scale, Retell’s combination of production benchmarking, engineered redundancy, and an emphasis on operational metrics illustrates a practical path to high-volume deployment. By relying on OpenAI models for the majority of live responses and using Codex to accelerate engineering, Retell aims to balance automation, reliability and human oversight as it grows its customer base and feature set.
Source: Read the original source
Leave a Reply