OpenAI has released three new models in its Realtime API that together push voice agents past the demo threshold and into production-grade enterprise deployments. The headline model is GPT-Realtime-2, which embeds GPT-5 class reasoning into a streaming voice runtime with end-to-end latency low enough for natural conversation. Alongside it, GPT-Realtime-Translate handles real-time speech translation across more than 70 input languages and 13 output languages, and GPT-Realtime-Whisper provides streaming transcription with the same latency profile as the conversational models.
The reasoning upgrade is the most significant of the three. Until this release, voice models in production had been substantially behind frontier text models in their ability to handle multi-step reasoning, edge case handling, and ambiguous user requests. The result was that voice agents performed well on scripted workflows like appointment booking and FAQ answering, but degraded quickly on conversations requiring inference, judgment, or recovery from misunderstood inputs. GPT-Realtime-2 closes that gap by exposing GPT-5’s reasoning capability through the streaming API, making it possible to build voice agents that can handle the kind of complex calls that previously required a human operator.
The two reference customers OpenAI has named illustrate the use case range. Zillow has deployed GPT-Realtime-2 for client calls covering home valuations, financing scenarios, and listing strategy questions, the kind of conversation where the customer expects a knowledgeable counterparty rather than a script-bound responder. Deutsche Telekom has deployed GPT-Realtime-Translate for multilingual customer support across European markets, allowing a single call center workforce to handle inquiries in dozens of languages without language-segmented staffing. Both deployments would have been impractical with the previous generation of voice models.
The economic implications for the call center industry are difficult to overstate. Traditional call center operations rely on labor arbitrage, with most large enterprises routing voice support to staff in the Philippines, India, Mexico, or Colombia. Those operations remain economical because human agents can handle conversational ambiguity that automated systems cannot. With GPT-Realtime-2, the conversational gap narrows substantially, and the per-minute economics of AI voice inference are an order of magnitude below human labor cost. The conclusion that traditional call centers operating in 2027 will do so by inertia rather than by economic logic is not hyperbole; it follows directly from the cost curves.
For sales teams tired of cold leads, slow customer responses, and manual processes, Dapta is the ultimate tool.
Dapta is the leading platform for creating AI sales agents specifically designed to increase inbound lead conversion. Respond to your leads in less than a minute with voice AI and WhatsApp that converts.
If you want your team to sell more while AI handles the complex stuff, you have to try it.
The translation model is a second-order disruption to a different industry. Real-time interpretation services for international meetings, live events, and customer support have historically required certified human interpreters at hundreds of dollars per hour. GPT-Realtime-Translate operates at a fraction of that cost with quality that, while imperfect, is sufficient for the majority of business communication needs. The simultaneous interpretation industry will experience a contraction similar to what document translation experienced after Google Translate and DeepL became production-quality in the 2010s.
For LATAM markets, the implications cut in two directions. On one hand, the region’s large call center sector faces a competitive threat from AI voice agents, with Mexican and Colombian operations particularly exposed because they have built their value proposition around English and Spanish bilingual support. On the other hand, LATAM enterprises that adopt AI voice agents early can leapfrog the call center expansion phase entirely, building customer service operations that scale to hundreds of thousands of monthly conversations without proportional labor growth. The same applies to multilingual sales motions, which have historically required dedicated regional teams and can now be served by a single AI runtime that switches languages mid-conversation.
The broader pattern is that voice is becoming a commodity layer for AI applications, in the same way that text became a commodity layer in 2023. Customer service, sales, scheduling, qualification, and outbound prospecting are all conversation-heavy workflows that have remained labor-intensive precisely because automated voice was not good enough. With GPT-Realtime-2 in production, the labor argument no longer holds at the level of conversation quality. What remains are integration questions about how to connect voice agents to existing CRM systems, knowledge bases, and compliance workflows, which is a tractable engineering problem rather than a fundamental capability gap.
The companies that move first will define the operational playbook for AI-native customer service organizations. The companies that wait will find themselves competing against operators with structurally lower cost of service and structurally faster response times, which is a difficult position to defend.
