OpenAI has released a new generation of voice models that fundamentally change what voice applications can do. Rather than simply converting speech to text or playing back responses, these models can reason through complex requests, translate between languages in real time, and take action while conversations unfold. The three new models, available through OpenAI's API, represent a significant leap in making voice interfaces feel natural and capable.

What Are the Three New Voice Models OpenAI Released?

OpenAI introduced three distinct voice models designed for different aspects of voice interaction. GPT-Realtime-2 is the flagship model, built with GPT-5-class reasoning capabilities that allow it to handle complex requests and keep conversations moving naturally. The model can call multiple tools simultaneously, adjust its tone based on context, and recover gracefully when users interrupt or change direction.

GPT-Realtime-Translate focuses specifically on live multilingual conversations. It supports more than 70 input languages and 13 output languages, allowing each person in a conversation to speak in their preferred language while hearing real-time translations. This capability opens possibilities for customer support, cross-border sales, education, and global events.

GPT-Realtime-Whisper is a streaming speech-to-text model that transcribes speech live as speakers talk, providing immediate written records of conversations without waiting for pauses or sentence completion.

How Much Better Is GPT-Realtime-2 Compared to Previous Models?

The performance improvements are measurable and significant. On Big Bench Audio, a benchmark that evaluates challenging reasoning capabilities in audio systems, GPT-Realtime-2 with high reasoning effort scores 15.2% higher on audio intelligence than its predecessor, GPT-Realtime-1.5. When using the highest reasoning setting, GPT-Realtime-2 scores 13.8% higher on Audio MultiChallenge, which measures instruction following and context management in spoken dialogue.

Real-world testing with Zillow, a major real estate platform, demonstrated even more dramatic improvements. After prompt optimization, Zillow reported a 26-point lift in call success rate on their hardest adversarial benchmark, moving from 69% to 95% success. The company also noted that GPT-Realtime-2 showed material improvements on Fair Housing compliance, which is critical for their business.

What New Capabilities Make These Voice Models Different?

These models introduce several features that weren't available in previous voice systems. Developers can now enable preambles, short phrases like "let me check that" or "one moment while I look into it," so users know the agent is actively working on their request. The models can also call multiple tools at once and make those actions audible with phrases like "checking your calendar" or "looking that up now," keeping users engaged while tasks complete in the background.

Context handling has expanded dramatically. The context window, which determines how much conversation history the model can remember, increased from 32,000 tokens to 128,000 tokens, roughly equivalent to processing 100,000 words at once. This allows for longer, more coherent sessions and more complex task flows without losing track of earlier details.

• Adjustable Reasoning Levels: Developers can select from minimal, low, medium, high, and xhigh reasoning settings, with low as the default, balancing faster responses for straightforward questions with deeper thinking for complex requests
• Stronger Domain Understanding: The models better retain specialized terminology, proper nouns, healthcare terms, and other vocabulary that matters in production settings where accuracy is critical
• Tone and Delivery Control: The models can adjust their speaking style to match the moment, speaking calmly while resolving issues, empathetically when users are frustrated, or upbeat when confirming successful actions
• Recovery Behavior: When something goes wrong, the models can recover gracefully by saying things like "I'm having trouble with that right now," instead of failing silently or breaking the conversation entirely

How Are Companies Already Using These Voice Models?

Several major companies are building production applications with these new capabilities. Zillow is developing an assistant that can listen to requests like "find me homes within my BuyAbility, avoid busy streets, and schedule a tour for Saturday," then reason through the request, use tools to search listings, and complete the booking.

Deutsche Telekom is testing GPT-Realtime-Translate for multilingual voice support, where customers can speak in their preferred language and the system translates the conversation in real time while preserving meaning and keeping pace with natural speech patterns.

Priceline is working toward a future where travelers can manage entire trips by voice, searching for flights and hotels conversationally, handling changes like adjusting a hotel reservation after a flight delay, getting real-time updates on Transportation Security Administration wait times, and translating conversations once travelers are on the ground.

What Three Patterns Are Emerging in Voice AI Applications?

OpenAI identified three distinct patterns developers are building around with voice AI. Voice-to-action systems let people describe what they need and the system reasons through the request, uses tools, and completes the task. Systems-to-voice applications turn context into live spoken guidance, like a travel app proactively telling a traveler about flight delays and routing options. Voice-to-voice systems help live conversations continue across languages, tasks, or changing context, enabling real-time translation and multilingual support.

These patterns can work together in sophisticated applications. A single voice agent might combine all three approaches, understanding user requests, translating between languages, and taking action across multiple systems simultaneously. This convergence is what makes voice interfaces feel like genuine assistants rather than simple command-response systems.

How to Build Voice Applications With These New Models

• Start with GPT-Realtime-2: Begin by integrating GPT-Realtime-2 into your application through OpenAI's API, using the default low reasoning setting for straightforward interactions and increasing reasoning effort only when handling complex requests that require deeper thinking
• Implement Tool Calling: Design your application to let the model call multiple tools simultaneously, such as checking calendars, searching databases, or updating records, while making those actions audible to users so they understand what the agent is doing
• Optimize for Your Domain: Train your prompts and system instructions to use domain-specific terminology and proper nouns relevant to your industry, whether that's healthcare, real estate, travel, or customer support, so the model retains specialized vocabulary accurately
• Test Tone and Recovery: Experiment with different tone settings and recovery phrases to match your brand voice and user expectations, ensuring the agent responds appropriately to frustration, success, or confusion
• Expand to Multilingual: If serving global audiences, integrate GPT-Realtime-Translate to enable customers to speak in their preferred language while the system translates conversations in real time, preserving meaning across 70+ input languages and 13 output languages

The release of these voice models marks a shift in how developers approach voice applications. Rather than building simple speech-to-text systems, developers can now create voice agents that understand context, reason through complex requests, and take action in real time. The performance improvements and real-world success stories from companies like Zillow and Deutsche Telekom suggest that voice is becoming a genuinely capable interface for getting work done, not just a novelty input method.