GPT-Realtime
OpenAI's GA speech-to-speech model for the Realtime API, replacing the STT-LLM-TTS chain with one low-latency model.
About
GPT-Realtime went generally available on August 28, 2025 as the first production speech-to-speech model for OpenAI's Realtime API. Instead of chaining speech recognition, a language model, and text-to-speech, it takes streamed audio in and emits streamed audio out in a single model, removing the latency tax of handoffs and letting one model control turn-taking.
On OpenAI's benchmarks the GA model is a clear step up from the December 2024 preview: 82.8% versus 65.6% on Big Bench Audio reasoning, and 66.5% versus 49.7% on the ComplexFuncBench function-calling eval. It ships with image input, reusable prompts, and remote MCP server support, and connects over WebSocket, WebRTC, or SIP. Pricing launched 20% below the gpt-4o-realtime-preview it replaced.
The honest caveats: OpenAI has since made gpt-realtime-1.5 and gpt-realtime-2 the recommended defaults, so this entry covers the original GA snapshot (gpt-realtime-2025-08-28), and the benchmark figures above are vendor-reported. It is API-only with no open weights.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 87; Trending score 88/100 from researcher assessment.