Retell Grows to 100 Million Calls a Month on OpenAI Models and Codex

Retell AI, a North American startup, builds voice agents for contact centers. It now serves more than 4000 companies, and monthly traffic has grown from 30 million calls at the start of 2026 to more than 100 million by September. OpenAI models generate responses in 80% of calls, total consumption has exceeded 1 trillion tokens, and customers' resolution rate is above 80%. Revenue per employee stands at $1.6 million on an annualized basis. OpenAI published this case study on October 27, 2026.

How Retell picks a model for a live call

In an enterprise setting, a voice agent has to follow company policies, stay within its role and keep the thread when a conversation drifts off topic. To compare models, Retell built an internal benchmark, vCX: dozens of models are run through complex multi-turn dialogues taken from real production scenarios. The benchmark tests instruction following, grounding in the information provided, correct tool calling and natural-sounding speech.

In the October 13, 2026 run, GPT-6 Astra ranked first in instruction following with 89.7%. Earlier measurements, taken before Astra's release, showed a different picture: with reasoning enabled, GPT-5.5 led in the share of hallucination-free responses (88.0%) and in tool-calling correctness (88.3%), ahead of Gemini-3.1-Pro and Claude-4.6. Without reasoning, it also stayed first in the share of hallucination-free responses. Retell integrates new models on the day of release and reruns the benchmark.

What happens inside a call

Retell's orchestration layer runs the audio models, tracks the flow of the conversation and decides when the agent should speak, listen or react to an interruption. OpenAI models generate the response text and the tool-call requests. Retell then executes the calls through integrations with the customer's systems and can write the result to the customer record. If a human is needed, the agent briefs the operator on the context, answers the operator's follow-up questions and transfers the caller only after the operator confirms readiness. After the call, GPT-5.6 Terra analyzes the transcript and extracts outcomes and structured data.

On latency, the picture is this. Daytime traffic exceeds the nighttime minimum by more than 50 times, so Retell uses dedicated OpenAI capacity shared across all of its customers. Regional API endpoints shorten the network path and help meet data residency requirements in healthcare and finance. Retell also maintains its own telephony stack with a backup, as well as fallback paths for voice and model services.

• GPT-6 Astra — first place in instruction following (89.7%), median first-phrase latency of 883 ms, also used for automated UI testing
• GPT-6 Sol — median latency of 883 ms
• GPT-6.1 Sol — response generation and tool calls in calls
• GPT-5.6 Luna — best latency in the benchmark, 497 ms
• GPT-5.5 — 88.0% hallucination-free, 88.3% correct tool calls
• GPT-5.6 Terra — post-call transcript analysis
• OpenAI Realtime API and text-to-speech — available to customers as a voice engine
• Anker — 80.4% of inquiries resolved, NPS of 63, first-line support in the US and UK
• Pine Park Health — 55.7% of patient contacts ending in an appointment booking, 2.1 full-time positions of freed-up medical assistant time

Codex as a virtual colleague

Retell's economics rest on more than the models inside calls. Engineers give Codex a general requirement, and it studies the code, checks it against real data and explains what to change. One engineer said Codex does this analysis better and noticeably faster than he can. During implementation, Codex often spots details the developer missed, and during debugging it matches logs and database records against the code. The team also uses it in customer work, for example to write instructions for models in customers' applications faster.

Frontend testing is a separate story. When an engineer opens a pull request, GPT-6 Astra uses computer use to open the interface, click through it, run scenarios, try to break them and report the problems it finds. According to Retell, earlier models did this unreliably. The team continues to hire, but it needs fewer engineers than the business is growing.

What to take from the case

For anyone building voice agents, the case offers several practical lessons. You need your own benchmark built on real dialogues: at Retell, it decides which model goes into production. Different models belong on different tasks: fast Luna for latency, Astra for instruction discipline, Terra for post-processing. You need regional endpoints and a fallback at every layer. Finally, when peak load is 50 times the minimum, dedicated capacity is not a luxury but a condition for stable responses.

It is worth remembering that all the figures come from Retell itself and were published by OpenAI as a customer story. The vCX benchmark is internal and has no independent verification. The company's plans: expand quality control and simulations, add languages and enable auto-training from a knowledge base.