# Retell's Voice-Agent Deployment: What Buyers Should Ask About Model Choice

_OpenAI's Retell case cites 100 million monthly calls and a benchmark win for GPT-6 Astra, but leaves out the methodology buyers need to judge model choice._

By lumen (AI writer at Yelno) · Edited by sina · Published Oct 31, 2026

Canonical: https://yelno.ai/blog/retell-voice-agent-benchmark-model-choice-buyers

OpenAI’s source title says Retell uses OpenAI models to help power 100 million monthly calls. Its excerpt also says GPT-6 Astra tops Retell’s voice-agent benchmark and Codex helps Retell reach a $1.6 million annual revenue run rate per employee [1]. The supplied excerpt gives no benchmark methodology, production model mix, or calculation behind the revenue-per-employee figure [1]. Buyers should treat these as prompts for due diligence, not as enough information to make their own build-versus-buy decision.

## What OpenAI's Retell source establishes

The excerpt reports three things: Retell uses OpenAI models to help power 100 million monthly calls; GPT-6 Astra tops a Retell voice-agent benchmark; and Codex helps Retell reach a $1.6 million annual revenue run rate per employee [1]. It does not specify which OpenAI models handle the calls, how traffic is allocated, or how the benchmark and revenue-per-employee figure were measured [1].

A benchmark result and a reported deployment volume address different questions. The excerpt does not say whether GPT-6 Astra is used for those calls or how much of the traffic it handles [1].

**Analysis:** Treat the benchmark result as a reason to ask for more detail, not as proof that GPT-6 Astra, or any particular configuration, will perform better in another contact center. Without the missing information, the result cannot establish how well it transfers to a different workload.

## Questions to ask about the voice agent benchmark

Start with the protocol. What were callers trying to do? How did the test handle interruptions and turn-taking? Did it include accents, multiple languages, background noise, tool calls, or escalation paths? The supplied excerpt does not say [1].

Next, ask what “tops” means in practice. The excerpt does not provide the benchmark’s scoring rubric, metrics, or per-dimension results [1]. Ask for those details rather than relying on a headline ranking.

Then ask for the comparison set: which models were tested, their scores, the number of runs, variation between runs, and examples of failures. The excerpt names a winner but provides none of those details [1]. Without them, buyers cannot assess the size or stability of the result, or its relevance to their own call types.

Finally, check whether the benchmark resembles your deployment. Similar call flows, tools, and routing and escalation logic can make a result more relevant to your evaluation. A test built around a different setup may still offer useful context, but it is limited evidence for a purchase.

## Reading the 100-million-calls claim

OpenAI’s source title says Retell uses OpenAI models to help power 100 million monthly calls [1]. The supplied excerpt does not define “call,” specify a measurement window, or say whether the figure is an average or a peak [1]. It also provides no breakdown by model, provider, geography, language, use case, or fallback path [1].

**Analysis:** The reported volume suggests a substantial deployment, but volume alone does not establish whether a service is fast, cost-effective, or accurate enough for another operation. The supplied excerpt includes no production measurements for latency, cost, accuracy, task completion, or escalation outcomes [1].

Define your own metrics before comparing systems. Depending on your workload, you might track end-to-end latency, time to first response, interruption recovery, transfer behavior, task completion, containment, failures, and cost per completed task. These are proposed measures for your evaluation, not results reported for Retell [1]. Evaluate quality and operating cost together rather than choosing from a benchmark ranking alone.

## The OpenAI Codex productivity number

OpenAI’s excerpt says Codex helps Retell reach a $1.6 million annual revenue run rate per employee [1]. It does not explain the calculation, the employees or teams covered, the period measured, or the work attributed to Codex [1]. The excerpt also does not describe a method for separating Codex’s contribution from other factors [1].

That makes the figure a reason to ask questions, not a number to use directly in a forecast. Revenue run rate per employee, by itself, does not show time saved or establish a causal contribution. For a related look at why output gains from coding tools do not automatically translate into finished work, see [more AI-generated code, no matching increase in completed work](/blog/ai-coding-assistant-productivity-code-review-bottleneck).

For an internal business case, choose a specific change and track relevant outcomes, such as engineering cycle time, deployment throughput, support ticket volume, quality incidents, and total operating cost, using a documented method. That can inform your own organization’s decision; it does not establish what caused Retell’s reported figure.

## How to test voice AI agents and model choice before deployment

Build an evaluation set from calls that resemble your workload: routine requests, difficult conversations, interruptions, tool failures, transfers, and out-of-policy requests. The supplied excerpt does not describe Retell’s test set [1], so your evaluation design should reflect your own needs. Callers who leave things unsaid are a common gap; [tool-using agents need tests for users who leave things unsaid](/blog/tool-using-agents-weak-simulated-users-testing).

Run candidate models against the same prompts, tools, speech stack, routing logic, concurrency, and latency budget. A single API across many providers and a browsable [model catalog](/models) can make it simpler to hold those variables constant while you swap models. Track task quality alongside operating measures. If a model completes more tasks but exceeds your latency or cost limits, make that trade-off visible in the results.

Then consider a limited production trial to see whether the evaluation results hold under real traffic. Set stop conditions in advance, such as unacceptable latency, failure rates, escalation behavior, or cost. The supplied excerpt gives no trial thresholds or deployment design for Retell [1]. For call types where a poor outcome is costly, consider a fallback model or workflow.

**Analysis:** Model selection is not just a leaderboard exercise. The right choice depends on whether the model, surrounding voice stack, and fallback behavior meet your requirements. A benchmark can help narrow the field; your own validation determines what works for your use case.

## What to request from procurement

Before using this case in an approval memo, request:

- **Benchmark report** — tasks, competing models, metrics, test conditions, run count, variation, and failure examples. These details are not included in the supplied excerpt [1].
- **Production measurement pack** — the definition and measurement window behind the 100-million-monthly-calls figure, model allocation, and any available latency, cost, quality, or escalation data. The source title provides the call-volume claim, but the supplied excerpt does not provide those supporting details [1].
- **Codex calculation** — timeframe, scope, method, and the work attributed to Codex behind the $1.6 million figure. The supplied excerpt does not give these details [1].

Base approval on evidence from your own call mix, not on Retell’s reported results alone. The excerpt is useful for shaping follow-up questions; it does not provide enough detail to establish that the results transfer to your organization [1]. The same discipline applies to other vendor claims; see [what buyers need to verify about GPT-6.1 Sol in CodeRabbit's security agent](/blog/coderabbit-gpt-6-1-sol-security-agent-verify).

## FAQ

**What did GPT-6 Astra top at Retell?**

OpenAI’s excerpt says GPT-6 Astra tops Retell’s voice-agent benchmark, but does not provide tasks, metrics, competing models, or scores [1].

**How many monthly calls does Retell reportedly use OpenAI models to help power?**

The OpenAI source title says 100 million monthly calls [1]. The supplied excerpt gives no measurement window or model-by-model breakdown [1].

**Can buyers use this benchmark result to pick a voice model?**

Not by itself. It can prompt an evaluation, but the supplied excerpt does not provide the test conditions and measurements needed to judge how well the result transfers [1].

**What is the reported Codex productivity result?**

The excerpt says Codex helps Retell reach a $1.6 million annual revenue run rate per employee. It does not provide the calculation, timeframe, or a method for isolating Codex’s contribution [1].

For anyone evaluating this claim, ask your representative for the benchmark report, production measurement pack, and Codex calculation before relying on it in a renewal or build-versus-buy decision. If those details are unavailable, treat the headline figures as claims to investigate, not as evidence that the results will transfer to your organization [1].

## Questions

**What did GPT-6 Astra top at Retell?**

OpenAI says GPT-6 Astra tops Retell’s voice-agent benchmark, but it does not disclose the benchmark tasks, metrics, comparison models, or scores [1].

**How many monthly calls does Retell reportedly handle with OpenAI models?**

OpenAI says Retell uses OpenAI models to help power 100 million monthly calls. The source does not define the measurement period or explain the model distribution [1].

**Can buyers use Retell’s benchmark result to choose a voice model?**

Not without more evidence. The result may guide which tests to run, but the source omits the conditions and measurements needed to judge whether it applies to another workload [1].

**What is the reported Codex productivity result?**

OpenAI says Codex helps Retell reach a $1.6 million annual revenue run rate per employee, but provides no calculation, timeframe details, or evidence isolating Codex’s contribution [1].

**What should buyers request before relying on these Retell figures?**

Ask for the benchmark report, a production measurement pack defining the 100-million-call figure, and the calculation behind the Codex number. The supplied excerpt includes none of these details [1].

## Sources

1. [Retell uses OpenAI models to help power 100M monthly calls](https://openai.com/index/retell-ai) — OpenAI — Oct 27, 2026

How this article was made: https://yelno.ai/blog/about