OpenAI has introduced a new model designated `gpt-realtime` and announced updates to its Realtime API, according to a company blog post published 28 August 2025 [source]. The announcement describes the model as optimised for low-latency voice and audio interactions, with native support for speech-to-speech workflows that bypass intermediate text transcription.

The Realtime API, first released in beta in October 2024, now supports the new model alongside existing `gpt-4o-realtime` variants. OpenAI states that `gpt-realtime` delivers faster response times and reduced audio processing overhead compared to prior implementations. The model is available through the same WebSocket-based API used for earlier realtime endpoints.

Key changes include updated pricing structures for audio input and output tokens, revised rate limits for enterprise customers, and expanded language support for non-English voice interactions. OpenAI has also published revised documentation covering session configuration, function calling within voice contexts, and handling of interruptions during multi-turn conversations.

The announcement does not specify whether `gpt-realtime` shares the same underlying architecture as GPT-4o or represents a distinct model family. No benchmarks comparing accuracy, latency, or hallucination rates between `gpt-realtime` and predecessor models were provided in the blog post.

Developers using the Realtime API are advised to review updated API parameters and test existing integrations for compatibility. OpenAI has indicated that `gpt-4o-realtime-preview` models will remain available during a transition period, though no end-of-life date has been announced.

This marks OpenAI's third significant model release in 2025, following updates to GPT-4 Turbo in January and the introduction of GPT-4o variants in March.