Evidence at a glance
A Hundred Million Calls Point to an Operating System
OpenAI published a customer case study about Retell AI on October 27, 2026. Retell is a voice AI platform for contact centers, where businesses can build agents that answer questions, take action, and transfer callers to a person when needed. OpenAI says Retell serves more than 4,000 businesses, with monthly call volume rising from 30 million early in 2026 to more than 100 million in September.
That growth invites an obvious question: which model made phone agents better? The case points to a broader change, in which model capability is embedded in a real-time operating workflow. OpenAI models generate responses for about 80% of Retell calls, while customer resolution rates are reported as above 80%. The first figure does not mean OpenAI handles every call on the platform, and the second comes without a consistent published definition. Both numbers need to be read with those limits in mind.
The Model Proposes Actions; the Platform Makes Them Happen
A single model does not independently carry a call from greeting to resolution. Retell’s platform runs the audio models, tracks the conversation, and manages when the agent speaks, listens, or handles an interruption. OpenAI models use the business instructions and call context to generate responses or request tool calls. Retell then executes those calls through customer-system integrations and can write the outcomes back to customer records.
That makes natural-sounding answers only one part of reliability. The agent must also follow business rules, choose the right tools, return results to the relevant systems, and hand off cases that need human judgment. Retell says its agent briefs the representative with context during a transfer, then analyzes the transcript and extracts outcomes after the call. The product, in other words, is the coordination among the model, orchestration layer, customer systems, and human workflow—not just the model’s text.
Production Evaluations Are Closer to the Real Selection Problem
Retell built its vCX benchmark to compare models on multi-turn conversations, testing instruction following, grounding, tool use, and conversational naturalness. OpenAI says GPT-6 Astra scored 89.7% for instruction following in Retell’s October 13, 2026 evaluation, ranking first in that category. This supports a claim of leading performance on one measure, not a claim that the model led every category in the benchmark.
The case also cites earlier comparisons in which GPT-5.5, with reasoning enabled, recorded an 88.0% non-hallucination rate and 88.3% tool-call correctness, ahead of Gemini-3.1-Pro and Claude-4.6 on those measures. Results that vary by task are a reason not to rely on a single overall ranking. In a contact center, instruction following, tool use, latency, and the handoff experience can all matter. A benchmark helps narrow the choices, but it cannot replace validation in the actual business workflow.
What the Case Metrics Do—and Do Not—Show
Anker offers a concrete example: its tier-one support agents in the United States and United Kingdom handle order tracking, warranty checks, and troubleshooting. Anker reports an 80.4% case resolution rate and a Net Promoter Score of 63, and says the agents’ tone, pacing, and warmth met its expectations for human representatives. These figures suggest that a clearly bounded support workflow can be handled by an agent. They do not show that other industries or more complex support cases will produce the same results.
Pine Park Health uses agents for appointment reminders and proactive scheduling, with outcomes recorded in patients’ medical charts. Its team says the system freed medical-assistant time equivalent to 2.1 full-time roles, which was redirected to patient care. The examples show different kinds of operational value: resolving support cases in one setting and reallocating staff time in another. But both come from an OpenAI customer case study, with no independent audit described and no shared resolution-rate methodology. They should not be treated as universal performance benchmarks.
Scale and Internal Efficiency Still Need Accounting
As call volume grows, reliability extends beyond the model. The research material says daytime traffic on Retell’s platform can exceed the nighttime low by more than 50 times. Retell uses dedicated servers, regional API endpoints, and backup arrangements for telephony, voice, and model services. For a technical leader, these are not peripheral optimizations: a slowdown or failure in any part of the chain can turn model capability into queues, repeat calls, or added pressure on human agents. The available material does not report failover performance or latency distributions, so it is not enough to assess service levels.
Retell also uses Codex internally to explore unfamiliar code, build features, and debug issues, and says it uses GPT-6 Astra’s computer-use capability to test front-end flows. OpenAI pairs this account with an annual revenue run rate of $1.6 million per employee, but provides no employee count, calculation method, comparison group, or quantified contribution from Codex. The figure may indicate business scale; it does not demonstrate that Codex caused it. A practical approach for other companies is to start with a workflow that has clear rules, system write-back, and safe human handoff, then test resolution, latency, and failover on their own calls before expanding.