Retell AI processes 100 million voice agent calls per month and has reached $1.6 million in annual recurring revenue per employee, both figures driven directly by its adoption of OpenAI's GPT-4o and Codex models.
GPT-4o Astra now tops Retell's internal voice agent benchmark, beating competing models on latency and response quality metrics that matter in real-time telephony. Codex handles the code generation workload that keeps Retell's engineering headcount lean while output scales. The per-employee revenue figure is the number worth interrogating: it suggests AI tooling is compressing the labor denominator faster than most engineering orgs have modeled.
The original post details how Retell architected around OpenAI's API stack, which models they tested and rejected, and where latency bottlenecks still exist. If you build on voice infrastructure or are benchmarking LLMs for real-time audio applications, the implementation specifics are worth your time.
[READ ORIGINAL →]