TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Retell says its voice AI platform processed more than 100 million calls in September 2026, up from 30 million in early 2026. The company says OpenAI models generate responses for 80% of its calls and that customers achieve resolution rates above 80%; those figures come from Retell and OpenAI, not an independent audit described in the source.
Retell AI says its voice-agent platform handled more than 100 million calls in September 2026, up from 30 million in early 2026, with OpenAI models generating responses for 80% of calls. The figures, published in an OpenAI customer report, describe a fast-growing deployment of AI in contact-center operations, where automated agents answer questions, use business systems and transfer calls to human representatives when needed.
Retell says its platform now serves more than 4,000 businesses and that its customers achieve resolution rates above 80%. It also reports that usage across OpenAI models has passed 1 trillion tokens. The source does not define the calculation behind the resolution-rate figure or provide an independently audited count of calls, so these should be read as company-reported metrics.
The platform combines a voice-call system with models that generate responses and request actions through customer integrations. Retell’s software tracks the conversation and handles when an agent speaks, listens or responds to interruptions. Models including GPT-6.1 Sol, GPT-6 Astra and GPT-5.5 use the customer’s instructions and conversation context; Retell then executes requested tool calls through connected systems and can record outcomes in customer records.
Retell says human handoff is built into the process. When a call requires human judgment, the agent can brief a representative, answer follow-up questions and confirm readiness before connecting the caller. Afterward, Retell says it uses GPT-5.6 Terra to analyze transcripts and extract call outcomes and other data.
Scaling Automated Contact-Center Calls
The reported volume suggests that voice agents are being used for a substantial share of routine customer interactions, rather than remaining limited to small demonstrations. For businesses, the potential value is handling common requests at scale while reserving human staff for cases that need judgment or more complex assistance. Retell’s figures also show how much the service depends on model performance, call infrastructure and integrations working together.
Customer examples in the report illustrate both the possible gains and the limits of the evidence. Electronics company Anker reported an 80.4% case-resolution rate for Tier-1 post-sales support in the United States and United Kingdom, covering order tracking, warranty checks and troubleshooting. Pine Park Health said its agents recovered medical-assistant time equivalent to 2.1 full-time roles; its scheduling agent booked appointments in 55.7% of calls where it reached a patient. These are individual customer outcomes and do not establish that other deployments will achieve the same results.
The scale also puts reliability and privacy in focus. Retell says daytime traffic can exceed its overnight low by more than 50 times. If automated agents handle essential customer or patient interactions, service interruptions, incorrect responses and mishandled handoffs could affect both organizations and callers.
Products Worth Considering

Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green

Bose QuietComfort Wireless Noise Cancelling Headphones, Twilight Blue
As an affiliate, we earn on qualifying purchases.
How Retell Selects Voice Models
Retell says it compares models using its vCX benchmark, which tests difficult, multi-turn conversations based on production scenarios. The evaluations assess instruction following, grounding in supplied information, correct tool use and natural conversational language. In results dated October 13, 2026, Retell ranked GPT-6 Astra first for instruction following at 89.7%.
Retell also cited evaluations conducted before GPT-6 Astra launched. In those tests, it said GPT-5.5, with reasoning enabled, had the highest scores among models it compared for non-hallucination rate, at 88.0%, and tool-call correctness, at 88.3%. The company said GPT-5.5 also led on non-hallucination rate with reasoning turned off. These are Retell’s benchmark results; the source does not provide enough methodological detail to treat them as an independent, general ranking of all voice models.
OpenAI’s report says Retell makes new OpenAI models available on launch day and reruns its evaluations as models improve. Retell also offers OpenAI Realtime models through the Realtime API and OpenAI text-to-speech as a voice engine. The company’s model choices are therefore part of a broader system that includes its own call-management software, telephony and customer-specific configuration.
““We choose models the way our customers choose vendors: on real calls, not demos. OpenAI keeps earning its spot.””
— Bing Wu, Retell AI co-founder and CEO
Limits of the Reported Metrics
The source is an OpenAI customer report, and several central figures are supplied by OpenAI or Retell. It does not describe an independent audit of monthly calls, model usage, token consumption or customer resolution rates. It also does not specify how “resolution” is defined across different businesses, what period underlies the overall rate above 80%, or how many calls were included in each customer example.
The benchmark results are also company-reported. The material describes the test categories and gives selected scores, but does not provide a full account of the model versions, sample sizes, scoring procedures or all competitors’ results. That limits what readers can infer from the rankings beyond Retell’s own tests.
Other operational details remain unreported, including the share of calls transferred to people, the rate of incorrect or incomplete answers, and how the more than 100 million monthly calls are distributed among customers and use cases. The source says Retell uses regional API endpoints and backup systems, but gives no uptime figures or specific service incidents.
Retell’s Next Model and Scale Tests
Retell says it will continue to add new OpenAI models when they launch and rerun its benchmark as model capabilities change. That creates a recurring test for the company: whether new models can maintain instruction following, accurate tool use and responsive conversations under live traffic, rather than only in benchmark evaluations.
The company also says it uses dedicated servers shared across its customers to manage peak demand, OpenAI regional API endpoints to reduce network travel and support data-residency needs, and backup systems for telephony, voice and model services. Further reporting could clarify how these measures perform as call volumes grow, including published uptime, latency and handoff data.
For customers and callers, the next indicators to watch are whether Retell sustains its reported call volume and resolution rates, how often agents need human intervention, and whether results hold across different industries. The source does not announce a new product launch or provide a date for another performance update.
Key Questions
How many calls does Retell say it handles each month?
Retell says its platform exceeded 100 million calls in September 2026, compared with 30 million in early 2026. The figure appears in an OpenAI customer report and is not described as independently audited.
How much of Retell’s call volume uses OpenAI models?
According to the report, OpenAI models generate responses for 80% of calls on Retell. The source does not break that share down by model, customer or type of call.
Does Retell say its agents resolve more than 80% of requests?
Yes. Retell reports customer resolution rates above 80%. The report does not give a single definition or measurement period for that overall figure. It separately cites Anker’s 80.4% case-resolution rate for a specific support deployment.
What happens when a voice agent needs human help?
Retell says an agent can provide a representative with call context, respond to the representative’s follow-up questions and confirm the representative is ready before connecting the caller.
Are Retell’s model rankings independent evaluations?
No independent evaluation is described in the source. The benchmark scores and rankings are Retell-reported results from its vCX tests, including a 89.7% instruction-following score for GPT-6 Astra dated October 13, 2026.
Source: openai.com
