What OpenAI published
OpenAI’s customer story on Retell makes three claims, and they are worth separating because they are different kinds of evidence:
- Scale: Retell handles 100 million calls per month.
- Quality: GPT-6 Astra ranks first on Retell’s own voice-agent benchmark.
- Efficiency: the company reached a $1.6M annual revenue run rate per employee, with Codex part of how the engineering work gets done.
The first is an operational fact, the second is a vendor-run benchmark, and the third is a financial ratio. Reading them as one “AI works” claim is how people end up disappointed; reading them separately is how you learn something usable.
The benchmark is the interesting part
A voice agent is not judged on a single response. It is judged on whether the call completes: did it understand the interruption, did it recover from a misheard name, did it stay on policy when the caller got angry. That is a task-level metric, and it is the right shape.
Most agent evaluation still looks like model evaluation: a prompt, a reference answer, a similarity score. That measures the brain in isolation. Retell’s benchmark measures the whole loop, which is why a model can be “better” on a public leaderboard and still lose on the phone.
The caveat is obvious and should be stated: this is Retell’s benchmark, published through OpenAI’s marketing channel. It tells you how a serious voice company evaluates its own stack. It does not tell you Astra beats every alternative in your use case.
Latency is a product requirement, not a tuning detail
You cannot run 100M calls a month on a system where latency is an afterthought. In voice, dead air is a bug: 800 ms of silence reads as a dropped connection, and a 3-second pause reads as a broken product.
The same discipline applies to desktop agents, just with different thresholds. A click that takes four seconds looks broken even if the work is correct. A form that fills in visibly wrong order makes the user reach for the mouse. If your agent’s step latency is measured in tens of seconds, nobody will watch it, and an agent nobody watches is an agent nobody schedules.
What the revenue-per-employee figure implies
$1.6M annual revenue per employee is not a normal SaaS ratio. It implies a small team supporting a very large call volume, which only works if the routine parts of the operation are automated - and if the engineering team leverages models heavily in its own work, which is where Codex comes in.
For a builder, the transferable idea is the split: automate the volume, keep humans on the exceptions. Retell’s product does that for phone calls. A desktop agent should do the same for the repetitive parts of a desk job - the difference is that the exceptions arrive as a file you have to look at, not a caller who has to be transferred.
What to copy on a Windows desktop
- Measure the task, not the model. Define success as “the invoice was filed correctly”, not “the response looked right”. Then count task success rate over twenty runs.
- Set a latency budget per step. Pick a number you would tolerate while watching, and treat exceeding it as a failure mode worth fixing.
- Automate volume, escalate exceptions. The agent should stop and ask when a file does not match the pattern it expects - that handoff is the product, not a limitation.
- Use the model in your own build loop. If Codex is good enough to help a 100M-call company ship, it is good enough to write your test fixtures.
MOOGH applies the first three to desktop work: each task is scoped, each run is logged, and anything outside the pattern stops for approval instead of guessing. Download MOOGH to see the handoff.
Related reading
FAQ
Is 100M calls a month verified?
It is the figure OpenAI publishes in its customer story, attributed to Retell. There is no independent audit in the piece.
Does this mean voice agents are solved?
No. It means one vendor has a benchmark, a scale figure, and a workflow worth reading. Reliability on unusual calls - accents, background noise, multi-party conversations - is still where deployments fail.
Should a desktop agent borrow the voice playbook?
The evaluation and latency parts, yes. The audio-specific parts obviously not. The common core is: measure the completed task, budget the latency, and escalate instead of guessing.