GPT-6 Astra won Retell's benchmark, yet it is missing from Retell's recommended models

Retell, a voice agent platform for contact centers, handles more than 100 million calls a month, according to an OpenAI case study. In early 2026 the figure was 30 million. The study is a vendor advertisement, and it leaves out one useful detail: what Retell actually recommends to paying customers.

OpenAI presents Retell as a customer running 100 million calls a month

Retell serves more than 4,000 businesses. Companies define an agent's context, goals and rules, and developers add workflows and integrations with their own systems. During a call, Retell runs the audio models and decides when the agent speaks, listens or handles an interruption. According to the study, OpenAI models generate responses and tool calls on 80% of calls. Retell has consumed more than 1 trillion OpenAI tokens.

Retell picks models with its own vCX benchmark, which compares dozens of models on difficult multi-turn conversations drawn from production. In the October 13, 2026 run, GPT-6 Astra ranked first for instruction following at 89.7%. An earlier evaluation from before Astra launched put GPT-5.5 on top for both non-hallucination rate (88.0%) and tool-call correctness (88.3%) with reasoning enabled.

The second half of the study is about Codex. Retell's engineers use it to research unfamiliar code, build features and debug. GPT-6 Astra clicks through front-end pull requests via computer use and tries to break them. Retell reports $1.6 million in annual revenue run rate per employee.

For voice agents, latency and per-minute price matter more than first place

Retell's price list tells a different story from the case study headline. GPT 6 Astra costs $0.32 per minute of call time. The cheapest models marked as recommended, GPT 5.6 Terra and Claude 5 Sonnet, cost $0.064, one fifth as much. Astra is not on the recommended list. A competitor's Claude is.

That makes sense. A contact center pays by the minute, and callers hang up when the agent goes silent. In the same benchmark, GPT-6 Astra and GPT-6 Sol both recorded a median first-sentence latency of 883 ms, while the fastest model, GPT-5.6 Luna, managed 497 ms. Daytime traffic on the platform can reach more than 50 times the overnight low, so every tenth of a second and every cent gets multiplied.

For a product team choosing a voice agent, the practical lesson is simple. First place in instruction following tells you which model sticks to the rules best. It does not tell you which one pays off on a line with thousands of concurrent calls. Retell itself mixes models by task: GPT-6.1 Sol, GPT-6 Astra and GPT-5.5 during calls, and GPT-5.6 Terra to analyze transcripts afterwards.

The success and productivity figures come from Retell and lack independent checks

The resolution rate above 80% is a figure Retell reports for its customers. At Anker, which uses the agents for post-sales support in the US and UK, it was 80.4% of cases resolved with an NPS of 63. Pine Park Health recovered medical-assistant time equal to 2.1 full-time roles. The study does not describe the vCX methodology or define what counts as a resolved call.

The $1.6 million per employee figure is slipperier still. Retell has not disclosed its headcount, and the study does not separate what Codex contributed from what fast company growth did. It is a revenue metric, not a measure of engineering output. The benchmark is also run by a company whose commercial relationship with OpenAI is the reason the study exists.

Retell's next recommendation will settle the question

Retell says it reruns the benchmark for every new model and ships new OpenAI models on launch day. So the best signal will be its own price list. If GPT-6 Astra joins the recommended models, the case study was right. If the recommendations stay with cheap models from two vendors, contact centers are being run on cost and latency.

Also worth watching is support for more languages, which Retell plans to add. European buyers will care about OpenAI's regional API endpoints, which the study says help meet data residency requirements. The study does not say which regions Retell actually uses.

Lilith's verdict

Retell let GPT-6 Astra win the race on the test track, then fitted customer cars with engines at a fifth of the price. A contact center director paying for every minute of talk time will read the price list before the case study.