OpenAI Unveils GPT‑Realtime‑2, Translate, and Whisper with 128K Context and Parallel Tool Calls

On 6 May 2026, OpenAI rolled out three new real‑time voice models—GPT‑Realtime‑2, GPT‑Realtime‑Translate, and GPT‑Realtime‑Whisper—through its Realtime API. The flagship GPT‑Realtime‑2 model expands the context window to 128 K tokens and allows up to 32 K output tokens, quadrupling the limits of the earlier 4o‑based Realtime‑1.5. It also introduces simultaneous tool calls, letting the model invoke multiple external services in parallel while delivering audible preambles such as “checking your calendar.” Developers can fine‑tune the model’s reasoning effort across five levels (minimal, low, medium, high, xhigh) and adjust tone, domain understanding, and graceful‑recovery messages, giving them granular control over conversational behavior.

GPT‑Realtime‑Translate extends live speech translation to more than 70 input languages and 13 output languages, enabling real‑time multilingual dialogue. GPT‑Realtime‑Whisper streams transcription and captions in real time, prioritising usability over raw audio fidelity. The release emphasises natural, responsive voice agents that can handle interruptions, maintain longer conversations, and retain specialised vocabularies. OpenAI highlighted that the new models are immediately available in the Realtime API, though upgrades to the ChatGPT voice interface are still pending.

The launch follows the earlier Realtime‑1.5, which was limited to a 32 K context window and single tool calls. By offering a larger context, parallel tool usage, and finer‑grained control, OpenAI aims to make voice‑enabled applications more robust and adaptable across industries. The company urges developers to integrate the new models now to take advantage of the expanded capabilities, while the broader AI community watches for how these enhancements will shape the next generation of conversational agents.