Powered by OpenAI Models: Inside Retell's 100M Monthly Calls AI Voice Stack

Retell has announced a significant figure in the AI voice space: its platform handles 100 million calls per month, and the core supporting this scale is a model portfolio from OpenAI. According to the original report, GPT-6 Astra ranks first in Retell's voice agent benchmark, while the adoption of Codex has helped the company reach $1.6 million in annualized revenue per employee.

This set of information connects three things: the voice capabilities of the underlying models, engineering-side development efficiency, and business-side productivity per employee. For teams evaluating voice agent tech stacks, Retell's example offers a reference that can be broken down.

GPT-6 Astra Tops the Voice Agent Benchmark

Retell maintains its own voice agent benchmark to compare different models' performance in real call scenarios. In this test suite, GPT-6 Astra took first place.

The difference between voice agents and pure text conversation lies in the longer pipeline: it needs to handle interruptions, noise, speech rate changes, multi-turn context, and also complete the handoff between speech recognition, reasoning, and speech synthesis under low-latency constraints. The significance of a benchmark ranking is not a score in a single dimension, but rather the overall stability a model reflects in an end-to-end call flow. Retell uses this result as an important basis for choosing OpenAI models.

For developers, the value of such third-party benchmarks lies in providing a reference perspective beyond official evaluations—after all, failure modes in voice scenarios often do not show up on text benchmarks.

The end-to-end pipeline of a voice agent typically consists of three core components: automatic speech recognition (ASR) converts user audio to text, a large language model (LLM) handles semantic understanding and dialogue decisions, and text-to-speech (TTS) converts the response back to audio. Unlike pure text conversation, this pipeline requires all three to work together within a strict latency budget. The industry often uses "end-to-end response time" (from when the user finishes speaking to when they hear the response) as a key metric, ideally controlled within 1 second. In addition, interruption handling relies on voice activity detection (VAD) to determine when the user starts speaking, and multi-turn context requires the model to maintain state under streaming audio input. Any latency jitter or recognition error in these steps is directly exposed to the user, so failure modes in voice scenarios are far more complex than text benchmarks.

From Model to Scale: The Engineering Challenges Behind 100M Monthly Calls

100 million calls per month is a threshold of magnitude. It means the system cannot merely run a demo under low traffic; it must simultaneously hold up in concurrency, fault tolerance, and cost control.

The cost structure of voice agents differs markedly from text applications: real-time processing of audio streams is highly latency-sensitive, and any jitter in a model call will be directly heard by the user. Pushing call volume to the hundred-million level requires a large number of trade-offs in model selection, routing strategy, and infrastructure. Retell's approach is to bind deeply to OpenAI's model system rather than broadly diversify across multiple vendors. This reduces integration complexity, but it also means a higher dependence on a single model vendor—a classic trade-off between efficiency and risk at scale.

Model routing across multiple vendors is a common strategy for scaling AI applications: dynamically dispatching requests to different models based on task type, latency requirements, or cost budget. The advantage is that it allows selecting the best option on demand and diversifying vendor risk, but the cost is increased integration complexity—maintaining multiple APIs, a unified abstraction layer, and a cross-vendor evaluation system. Retell's choice to bind deeply to a single OpenAI system essentially internalizes routing logic within the same vendor's model portfolio (for example, using models of different specifications for reasoning and lightweight tasks respectively), reducing integration and operational overhead. This "efficiency-first" trade-off is often more worthwhile in the early stages of scaling, but it amplifies risk when vendor pricing changes or model iteration cadence shifts.

Codex and $1.6 Million Annualized Revenue per Employee

Another figure worth noting in the original report: Codex helped Retell reach an annualized revenue run rate of $1.6 million per employee.

This metric reflects how engineering efficiency translates into business results. The value of AI coding tools is usually discussed at the level of "writing code faster," but what truly affects a company's financial structure is whether it can enable a smaller team to support a larger product surface. For a company building voice infrastructure, code volume, integration complexity, and operational needs are all substantial, so productivity data becomes a direct entry point for measuring the actual return of AI-assisted development.

It should be noted that annualized revenue per employee is a composite metric affected by pricing, customer mix, and team size, and cannot be simply attributed to a single tool. But presenting it alongside the use of Codex at least shows that Retell internally recognizes this investment.

Annualized Revenue per Employee is a common efficiency metric in the SaaS industry, typically calculated as annual recurring revenue (ARR) or annualized revenue run rate divided by the number of full-time employees. For reference, mature SaaS companies generally have annualized revenue per employee in the range of $200,000 to $500,000, and top efficiency-focused companies can exceed $1 million. The $1.6 million level is in a relatively high range, but note that this metric is greatly affected by company stage, pricing strategy, and customer mix—early-stage companies with lean teams and a small revenue base tend to have inflated metrics. Therefore, it is better suited as a longitudinal trend observation than as an absolute yardstick for cross-company comparison.

The Logic of Choosing a Voice Agent Tech Stack

Retell's case reflects a reality in the current voice agent field: model capability remains the key variable determining the upper limit of product experience.

The margin for error in voice interaction is much smaller than in text, and users have very low tolerance for latency and interruption handling. This makes teams more inclined to choose models that have been validated at scale on voice pipelines, rather than simply comparing parameters or price. Retell uses Astra for voice agents and Codex for the development process, essentially configuring the "product-side model" and the "engineering-side model" separately, each matched to the most suitable task.

This division-of-labor approach is worth referencing for other AI application teams: not every step needs the same model, and splitting configuration by task characteristics is often more pragmatic than a unified tech stack.

Concerns After Scaling

The hundred-million-level call volume and high productivity data form a beautiful story, but there are parts that need to be viewed calmly.

The benchmark ranking is maintained by Retell itself, and the test set and scoring criteria were not disclosed in the source material, making external reproduction difficult. Deep reliance on a single model vendor leaves it passive when pricing changes or model iteration cadence shifts. In addition, competition in the voice agent space is intensifying, and the lead window in model capability is usually not very long.

Retell's current advantage is built on the combination of "model capability + engineering efficiency." Whether this combination can hold as call volume continues to grow depends on whether it can convert scale into structural barriers in data and cost, rather than merely resting on having chosen a stronger model.