OpenAI Realtime API
Speech-to-speech voice agents on gpt-realtime-2.1, billed per audio token
Overview
The OpenAI Realtime API streams audio in and out of a single speech-to-speech model over WebRTC, WebSocket or SIP, with interruption handling, tool calls and remote MCP servers. It launched in beta in October 2024, went generally available in August 2025, and runs on gpt-realtime-2.1 and gpt-realtime-2.1-mini since July 2026.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Cached audio input $0.40 per 1M
- Text $4 input, $0.40 cached, $24 output per 1M
- Image input $5 per 1M ($0.50 cached)
- Cached audio input $0.30 per 1M
- Text $0.60 input, $0.06 cached, $2.40 output per 1M
- Image input $0.80 per 1M ($0.08 cached)
- Billed per second, not rounded up to a minute
- Backend model and tool usage charged separately
API pricing (pay as you go)
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-realtime-2.1 (audio) | $32 | $0.40 | $64 |
| gpt-realtime-2.1 (text) | $4 | $0.40 | $24 |
| gpt-realtime-2.1 (image) Image input only. | $5 | $0.50 | — |
| gpt-realtime-2.1-mini (audio) | $10 | $0.30 | $20 |
| gpt-realtime-2.1-mini (text) | $0.60 | $0.06 | $2.40 |
| gpt-realtime-2.1-mini (image) Image input only. | $0.80 | $0.08 | — |
Use cases
What you can produce with OpenAI Realtime API
- Speech-to-speech voice agents from a single model, with no separate STT or TTS stage
- Browser connections over WebRTC and server connections over WebSocket
- Phone calls answered through a SIP trunk connected directly to OpenAI
- Function calling and remote MCP tools during a live voice session
- Barge-in handling when a caller interrupts the agent
- Image input alongside audio and text in the same session
- Cached audio input billed at $0.40 per 1M tokens on gpt-realtime-2.1
The Realtime API removes the speech-to-text and text-to-speech hand-offs that made earlier voice bots slow, and since the August 2025 GA it handles phone calls over SIP, MCP tools and image input. gpt-realtime-2.1 adds configurable reasoning; the mini model costs a third as much for audio input.
Audio tokens are pricey, so cost depends heavily on how much the agent talks, and you are limited to OpenAI's voices. OpenAI's newer GPT-Live, billed per minute, is now an alternative worth testing for full-duplex conversation.
What we found
The Realtime API is OpenAI's interface for speech-to-speech applications: one model listens and replies in audio over WebRTC, WebSocket or SIP, instead of a chain of speech-to-text, a text model and text-to-speech. OpenAI opened it in public beta in October 2024 and made it generally available on 28 August 2025, adding remote MCP servers, image input and phone calls over SIP, along with the gpt-realtime model and two new voices, Marin and Cedar. The beta interface was switched off on 12 May 2026.
The current models are gpt-realtime-2.1 and gpt-realtime-2.1-mini, announced on OpenAI's developer forum on 6 July 2026. The full model has configurable reasoning effort and improves on GPT-Realtime-2 at recognising letters and numbers, handling silence and background noise, and coping with interruptions; OpenAI said the release also cut p95 latency across its realtime voice models by at least 25%. Older families, including gpt-realtime, gpt-4o-realtime and gpt-realtime-mini, are scheduled for removal on 20 January 2027.
Billing is per token, and audio is the expensive part. gpt-realtime-2.1 costs $32 per million audio input tokens and $64 per million audio output tokens, with cached audio input at $0.40; text is $4 in and $24 out. The mini model charges $10 and $20 for audio and $0.60 and $2.40 for text. What a minute of conversation costs depends on how long the agent talks and how much context it carries, so measure it on real traffic before committing.
The bigger change sits beside it. On 8 July 2026 OpenAI launched GPT-Live, a full-duplex voice model that can listen while it speaks and hands harder work to a separate backend model. It is now listed in the API at $0.05 a minute plus backend usage, with its own SIP support and a guide for migrating Realtime apps. OpenAI still documents Realtime as the starting point for low-latency agents with in-session tool calls, but a new voice project should compare the two.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
About this listing
This entry was compiled from publicly available data including OpenAI Realtime API's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with OpenAI Realtime API unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to OpenAI Realtime API directly →
Spotted something out of date? Suggest an update →
Alternatives to OpenAI Realtime API
More in Video & Audio
