Retell runs 100M monthly calls on OpenAI models as GPT-6 Astra tops its voice benchmark

OpenAI published a customer story on October 27 describing how Retell, a voice agent platform, runs its call traffic on OpenAI models. According to the post, Retell now handles about 100 million calls a month, with GPT-6 Astra among the models carrying that load.

The benchmark claim is the most concrete technical detail: GPT-6 Astra ranks first on Retell's internal voice agent benchmark. Evaluations of this kind typically test multi-turn stability, interruption handling and tool-calling accuracy, so Retell's choice signals that production call distributions matter more to it than general-purpose leaderboards.

The other half of the story sits on the engineering side. OpenAI says Retell uses Codex in development and has reached $1.6 million in annualized revenue per employee, a figure that implies a small team is shipping and supporting product at a rate conventional SaaS companies struggle to match.

One caveat is structural. This is a customer story published by the model vendor, with metrics presented jointly by both companies and no third-party audit. It establishes that Retell runs OpenAI models at scale in production; it is not independent evidence of how those models compare with rivals.

Voice agents have been one of the fastest-moving AI categories of the past year. Customer support, outbound calling and appointment confirmation all demand low latency and natural turn-taking, and they bill naturally by call volume, which gives the category a clear commercial model.

For OpenAI, holding up a high-volume account like Retell points to where inference demand is heading. Voice is a token-dense, always-on workload, and once a customer moves its telephony stack onto a provider, switching costs rise along with call volume and the bespoke evaluation suite built around it.

Two things are worth watching. The first is whether Retell publishes its benchmark methodology, which would let outsiders judge the size of GPT-6 Astra's lead. The second is whether more voice agent companies disclose revenue-per-employee figures, which would make the organisational impact of coding agents like Codex measurable rather than anecdotal.