Retell uses OpenAI models to support 100 million monthly calls
OpenAI says Retell uses its models across 100 million monthly calls, while GPT-6 Astra and Codex support voice-agent testing and growth.
OpenAI says voice-agent company Retell uses its models to help power 100 million calls each month, highlighting the scale that conversational AI infrastructure can reach when deployed in production. The announcement also credits GPT-6 Astra with leading Retell’s voice-agent benchmark and says Codex helped the company reach a $1.6 million annual revenue run rate per employee.
The claims position Retell as an example of two developments AI builders are watching closely: voice systems moving beyond demonstrations into high-volume call workflows, and small, specialized companies using foundation models and coding tools to increase output. However, the available evidence comes from OpenAI’s own company profile, so the adoption, benchmark, and productivity figures should be treated as vendor-reported rather than independently verified.
The announcement
OpenAI’s official profile of Retell describes the company as using OpenAI models in voice-agent applications that collectively handle 100 million monthly calls. The material does not provide a fuller breakdown of those calls, including the share handled by OpenAI models, the industries involved, geographic distribution, average call length, or whether the figure represents connected calls, attempted calls, or another internal measure.
The announcement also names GPT-6 Astra as the top performer in Retell’s voice-agent benchmark. The source summary does not disclose the benchmark’s test set, scoring method, competing models, or the margin separating GPT-6 Astra from other systems. Those details matter because voice-agent performance depends on more than text quality. Response latency, interruption handling, speech recognition, tool execution, reliability, and the ability to follow business rules can all affect whether an agent works in a real call.
OpenAI additionally says Codex helped Retell reach a $1.6 million annual revenue run rate per employee. That is a notable operating metric, but the source does not explain how much of the result is attributable to Codex, how the company defines employee count, or whether the figure reflects a temporary run rate rather than booked annual revenue.
Evidence and limits
The strongest confirmed point in the source set is that OpenAI is publicly presenting Retell as a customer or user of its models and associating the company with 100 million monthly calls. A separate news listing carries the same headline, but the available text does not add independent reporting or additional verification.
That makes the story useful as a signal of OpenAI’s commercial narrative, but not as a complete audit of Retell’s business. The 100 million-call figure is an adoption claim attributed to OpenAI. The GPT-6 Astra result is a benchmark claim attributed to OpenAI and Retell’s evaluation process. The $1.6 million figure is a vendor-reported productivity and revenue-run-rate claim, not evidence that every company using Codex can produce comparable economics.
The missing context is particularly important for buyers comparing voice platforms. A call count alone does not reveal whether conversations were successful, whether users reached a human agent afterward, how often the system required intervention, or what the underlying inference and telephony costs were. Likewise, a benchmark win may not predict performance in regulated support, complex sales, appointment scheduling, or other workflows with different accuracy and safety requirements.
Why the metrics matter to AI builders
For builders, the reported scale suggests that voice agents are increasingly being evaluated as operational systems rather than isolated model demos. At 100 million monthly calls, issues such as monitoring, fallback behavior, escalation policies, prompt and tool versioning, and per-call economics become central product concerns. The model is only one part of the stack; telephony, orchestration, speech processing, customer data access, and auditability also shape the result.
Retell’s benchmark claim points to another practical issue: model selection for voice is likely to remain workload-specific. A system that performs well in a controlled voice-agent test may still need extensive tuning for interruptions, accents, domain terminology, long calls, or strict compliance requirements. Product teams should therefore ask to see evaluation definitions and production error rates, rather than treating a single leaderboard position as a general guarantee.
The Codex claim matters for founders because it links developer tooling to company operating leverage. If Codex helped Retell increase output per employee, the implication is not simply that coding became faster. Smaller teams may be able to maintain more integrations, ship more experiments, or respond more quickly to customer requirements. But the available evidence does not separate Codex’s contribution from hiring, architecture, capital, pricing, or other business factors.
Implications for enterprise AI
Enterprise buyers can read the announcement as evidence that voice automation is being discussed at meaningful production volume, but they should focus on reliability and governance before call volume. A deployment review should include escalation to human staff, handling of ambiguous requests, protection of personal information, call recording policies, incident response, and the ability to change models without disrupting customer-facing workflows.
For model providers, Retell is also a useful showcase of downstream usage. OpenAI can point to a concrete application category in which model quality, latency, and developer tooling combine into a visible business outcome. Competitors will likely face pressure to demonstrate comparable production workloads rather than relying only on general benchmark scores.
For Retell, the public figures may help establish credibility with customers and investors, but they also create expectations. As usage grows, the company will need to show how it measures successful calls, controls costs, maintains service quality, and handles failures at scale. None of those operational details is provided in the current source material.
What to watch next
The next useful signal would be independent detail on Retell’s 100 million monthly calls: what counts as a call, how many are completed successfully, and which workloads generate the volume. Evidence from Retell customers or third-party usage data would help distinguish broad adoption from an OpenAI marketing milestone.
The GPT-6 Astra claim also warrants a published benchmark methodology, including test cases, baseline models, latency measurements, interruption handling, and reproducibility information. For enterprise teams, production metrics such as transfer rates, hallucination rates, average resolution time, and cost per completed interaction would be more actionable than a headline ranking.
Finally, observers should look for clearer evidence behind the Codex productivity claim. Retell’s engineering practices, team size, revenue definition, and before-and-after measures would determine whether the reported $1.6 million annual revenue run rate per employee reflects durable operating leverage or a snapshot of a fast-growing company.
Creati.ai perspective
OpenAI’s Retell profile is a meaningful commercial signal, but its numbers are best understood as a case study offered by the model provider, not a fully verified market report. The scale claim suggests that voice agents can support substantial workloads, while the limited disclosure leaves open the harder questions about quality, cost, and safety.
For AI builders and enterprise buyers, the practical lesson is to evaluate the whole voice-agent system. Model benchmarks and developer productivity matter, but production success will depend on measurable call outcomes, dependable escalation, transparent economics, and controls that remain effective as usage expands.
