OpenAI has released gpt-realtime, its most advanced speech-to-speech model, alongside the general availability of the Realtime API. The updates aim to reduce latency, improve speech quality, and give developers stronger tools, such as MCP server support, image input, and Session Initiation Protocol (SIP) phone calling support, for building production-ready AI voice agents.
The combined Realtime API and gpt-realtime is designed to handle end-to-end speech processing within a single system, rather than chaining together separate speech-to-text and text-to-speech models. This architecture cuts response times while preserving nuance in delivery, a critical improvement for real-time agents where even small delays can break conversational flow.
The gpt-realtime was trained to produce higher-quality speech with more natural pacing and intonation, and to respond reliably to style instructions such as “speak empathetically” or “use a professional tone.” Two new synthetic voices, Cedar and Marin, are available, and existing voices have been updated for greater realism.
On comprehension benchmarks, gpt-realtime shows measurable improvements. It can track non-verbal cues, switch languages within a single sentence, and more accurately process alphanumeric sequences (such as phone numbers, VINs, etc) across languages, including Spanish, Chinese, Japanese, and French. Internal testing highlights this jump, with gpt-realtime reaching 82.8% accuracy on Big Bench Audio compared to 65.6% for the previous model. Instruction-following is also sharper, with MultiChallenge audio benchmark scores rising from 20.6% to 30.5%.
Function calling is another area of focus. The model now performs better at identifying relevant functions, calling them at the right time, and supplying the correct arguments. On ComplexFuncBench, accuracy rose to 66.5% from 49.7%. There were updates to asynchronous function calling, allowing the voice agent to continue the conversation while waiting for results, a feature with obvious value for customer support and transactional applications.
The Realtime API has been upgraded to align with production requirements. Developers can now connect remote MCP servers directly into a session, enabling tool calls without manual integration work. Image input is supported, allowing applications to ground conversations in visual context, such as screenshots or photos. SIP support makes it possible to integrate voice agents with existing telephony systems, including PBXs and desk phones. Reusable prompts simplify session management, while full EU data residency support addresses compliance concerns for European deployments.
According to the release notes, early enterprise partners are testing these capabilities in production-like scenarios. Zillow is piloting voice-driven home search, while T-Mobile is exploring customer service use cases where real-time adaptability is essential. Both companies highlight the shift from scripted automation to more flexible, domain-specific expertise delivered through AI agents.
OpenAI has also reinforced safeguards around deployment. The Realtime API incorporates classifiers that can terminate harmful conversations, and developers can add domain-specific guardrails via the Agents SDK. Preset voices in Realtime API are used to reduce impersonation risks.
Both gpt-realtime model and Realtime API are immediately available to all developers. To get started, developers can visit the Realtime API documentation and prompting guide, and test the new gpt-realtime demo in the Playground.
