Latent SpaceMay 8, 2026, 15:11 UTC+8

[AINews] GPT-Realtime-2, Translate, and Whisper: new SOTA realtime voice APIs

[AINews] GPT-Realtime-2, -Translate, and -Whisper: new SOTA realtime voice APIs

OpenAI released GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper in the Realtime API, with GPT-Realtime-2 expanding context from 32K to 128K and scoring 96.6% on Artificial Analysis Big Bench Audio.

Why it matters: HKR-H/K/R all pass: an OpenAI real-time voice API refresh, a 32K→128K context jump, and a 96.6% Big Bench Audio claim. Score stays at 86 because this is a major API update, not a flagship foundation-model release.

Read the original ↗Export Markdown

Take

OpenAI is aiming GPT-Realtime-2 at production voice agents, not voice demos. The hard numbers line up: context jumps from 32K to 128K, Artificial Analysis reports 96.6% on Big Bench Audio, time-to-first-audio is 1.12s at minimal reasoning and 2.33s at high reasoning, and pricing stays at $1.15/hour input plus $4.61/hour output. Capability up, latency still usable, price unchanged is an unusually clean product move.

The piece I buy most is audible tool behavior and recovery language. In support calls, scheduling, and IT helpdesks, “checking your calendar” and graceful failure beat a nicer voice. Scale AI’s instruction-retention jump from 36.7% to 70.8% APR hits the old failure mode directly. But don’t blur this into a ChatGPT Voice upgrade; Simon Willison’s caveat matters, because the consumer surface has not moved yet.

Source: Latent Space · latent.space
Published: May 8, 2026, 15:11 UTC+8 · May 8, 2026