OpenAI's GPT‑Realtime‑2 hits 128K context

OpenAI just upgraded its realtime voice stack — the part of the API that lets apps listen, speak, and act while a person is still talking. The headline feature is GPT‑Realtime‑2, a new speech-to-speech model with a 128,000-token context window and stronger reasoning. That sounds like a spec-sheet bump, but the real change is simpler: voice agents can now remember more, stay coherent longer, and do harder tasks without constantly losing the thread. ### What actually shipped? OpenAI released three models on May 7, 2026: GPT‑Realtime‑2 for low-latency voice agents, GPT‑Realtime‑Translate for live speech translation, and GPT‑Realtime‑Whisper for streaming transcription. The package is aimed at realtime apps — customer support, assistants, meetings, broadcasts, lessons, and anything else where waiting for batch processing ruins the experience. (openai.com) ### Why is 128K context a big deal? Because realtime voice systems usually fall apart in long sessions. If the model only has a limited working memory, it starts dropping names, preferences, earlier instructions, or the state of a task. GPT‑Realtime‑2 expands the realtime window from 32K to 128K tokens — 4x larger — which means longer calls, bigger system prompts, and more room to keep structured state without trimming so aggressively. (openai.com) ### What does “reasoning voice model” mean here? Basically, the model is not just transcribing speech and firing back a canned answer. OpenAI positions GPT‑Realtime‑2 as a voice model with GPT‑5‑class reasoning, configurable reasoning effort, stronger instruction following, and more reliable tool use. In practice, that means a voice agent can pause briefly, decide what to do, call a tool, and then continue the conversation without sounding like it switched brains halfway through. (developers.openai.com) ### How is that different from live translation? Translation is a separate lane. GPT‑Realtime‑Translate is built to act like an interpreter, not a conversational assistant. It streams translated audio and transcript updates while the speaker is still talking, supports more than 70 input languages, and currently outputs into 13 target languages. If you want an agent that answers questions and uses tools, OpenAI says to use GPT‑Realtime‑2 instead. (openai.com) ### Why does this matter for builders? Because the old pattern was messy. Developers often had to stitch together speech recognition, a text model, a translation layer, text-to-speech, and app logic — each with its own latency and failure modes. OpenAI’s pitch is that these newer realtime models collapse more of that stack into a single session over WebRTC, WebSocket, or SIP. Fewer handoffs usually means lower lag and fewer weird conversational breaks. (openai.com) ### What kinds of apps get better first? Anything with long, stateful spoken interaction. Think support calls where the agent has to remember an order number from 10 minutes ago, tutoring sessions that build on earlier mistakes, meeting copilots that can translate and summarize on the fly, or creator tools that generate captions and multilingual audio during a live stream. The bigger context window is the enabler — it keeps the session from feeling amnesiac. (platform.openai.com) ### What’s the catch? More reasoning is not free. OpenAI says higher reasoning effort can increase latency and output-token usage, so builders have to trade responsiveness against depth. That matters a lot in voice, where even a short awkward pause feels broken. So the win here is not “maximum intelligence at all times.” It’s finer control over when the model should think longer and when it should just answer fast. (openai.com) ### So what changed in the voice race? The shift is from voice as a thin interface to voice as a working agent. Realtime audio models used to be impressive mostly when the conversation was short and simple. GPT‑Realtime‑2 pushes that boundary outward — longer memory, better tool use, and more reliable instruction following in the same live session. For anyone building spoken software, that is the part that matters. (openai.com) (developers.openai.com)