GPT-Realtime-2.1
Released Jul 6, 2026
OpenAI's current Realtime API speech-to-speech reasoning model, with configurable reasoning effort and tool use. It updates GPT-Realtime-2 with better alphanumeric recognition, silence and noise handling, and interruption behavior.
- Price per minute
- —
- Pricing
- Per 1M tokens: text $4 in / $0.40 cached / $24 out; audio $32 in / $0.40 cached / $64 out; image $5 in / $0.50 cached.
- Inputs
- audio, text, image
- Tool calling
- Yes
- Languages
- —
- Voices
- 10
- Big Bench Audio
- 95.8%
- Access
- Generally available
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Configurable reasoning effort in a single speech-to-speech model
- Function calling and prompt caching on the Realtime API (WebRTC, WebSocket, SIP)
- Accepts image input alongside audio and text
- 128K context, 32K max output
Weaknesses
- Audio at $32 in / $64 out per 1M tokens, far above per-minute rivals
- Big Bench Audio drops from 95.8% at high effort to 86.8% at minimal (AA)
Notable: frontier-lab clause (also #8 on AA Speech to Speech Index, 75.2 at High). AA Big Bench Audio: 95.8% (High), 86.8% (Minimal). OpenAI's deprecations page names gpt-realtime-2.1 as the replacement for gpt-realtime and gpt-4o-realtime (shutdown 2027-01-20). Knowledge cutoff 2024-09-30. Voices: 10 built-in Realtime voices (alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin, cedar) per the Realtime conversations guide, and custom voices are available to eligible customers. No language count is published.
Links
API model strings
- openai
gpt-realtime-2.1