OpenAI Realtime API
Realtime speech-to-speech API for building low-latency voice agents over WebRTC, WebSocket, or SIP.
- Pricing at a glance
- $32/M audio input tokens and $64/M audio output tokens for gpt-realtime
- Starting price
- From $32/M tokens
- Setup
- Developer build
- Pricing model
- Usage Based
- Try it
- Public web demo
Use an eligible API account to examine real-time model interaction.
Your application's tools, controls and deployment still need testing.
At a glance
Calls & telephony
- DetailCalls
Inbound & outbound
- SupportedHuman handoff
Yes
- Not supportedPhone numbers included
No
- SupportedBring your own carrier
Yes
- SupportedSIP trunking
Yes
Buy number from a third-party SIP trunking provider, point trunk at OpenAI SIP endpoint
Trust & access
- With conditionsHIPAA
Enterprise plans
BAA plus Modified Retention provisioning required; Realtime eligibility unconfirmed
- SupportedSOC 2
Type II
- SupportedGDPR
Yes
DPA available for API customers (from search snippet; page returned 403)
- Not publishedLanguages
Not published
- Not publishedFree to start
Not published
- SupportedDeveloper API
Yes
Morak scorecard
How we score →- Ease of setupDeveloper build
Expect developers to build call logic, integrations and deployment.
- Pricing transparencyPlans published
Plan prices are listed on the vendor's site.
- Try before you buyWeb demo
A public web demo shows the experience, not your own setup.
- Call & telephony controlBroad
Documented: inbound, outbound, human handoff, bring-your-own carrier and SIP.
- Compliance readinessStrong
Published: SOC 2 Type II, GDPR and HIPAA on enterprise plans.
- Integration coverage2 of 4 covered
Connects to phone system and automation or API. Not listed: customer records (CRM, EHR or POS) and calendar.
Typical score among developer APIs
Use a real-time speech model inside your own app
What you get
Build spoken interactions over the supported real-time connections with custom tools.
What to account for
You still build application logic and operational controls. A framework such as LiveKit can complement the model API rather than replace it.
Best for
- Developer teams
- Speech-to-speech agents
- Tool-using voice workflows
- Custom web and phone experiences
Not a fit for
- Non-technical teams wanting a finished receptionist
- Teams needing bundled telephony and CRM setup
- Use cases without application-level safety controls
What the platform provides
- Model capabilities
- Audio and text input and output, instruction following and tool calls within a conversation.
- Connection options
- WebRTC, WebSocket and SIP support browser, app and phone-agent architectures.
Connections & compatibility
Telephony
- WebRTC
- SIP
AI models & voice
- OpenAI Agents SDK
Automation & APIs
- WebSocket
- Custom tools
- Custom API
Data access & permissions
What an agent built on OpenAI Realtime API can touch once deployed. These are the first questions a security review will ask.
- Reads customer dataYes
- Writes to your systemsYes
- Sends messagesYes
- Triggers workflowsYes
- Records callsNo
- Needs API keysYes
- Human approval stepYes
Use scoped API keys and server-side tool authorization. Disclose AI interaction, prohibit deceptive impersonation, obtain consent for recording or custom voices, and do not expose sensitive tools directly to model-generated arguments.
Pricing & terms
OpenAI lists gpt-realtime at $32 per million audio input tokens and $64 per million audio output tokens, with separate text-token rates. Telephony, storage, observability, and other services are additional. Verify the current model page before estimating per-minute cost.
Before committing
Test misheard names, spoken prompt injection and tool timeouts. Measure audio, text and infrastructure usage on the same pilot.
How it compares
Side by side with two alternatives buyers weigh against it, on the same facts and scorecard.
| Detail | OpenAI Realtime API This productRealtime speech-to-speech API for building low-latency voice agents over WebRTC, WebSocket, or SIP. | Works alongside it Open-source framework and managed cloud runtime for building and deploying realtime voice agents. Why compare: Consider LiveKit for the session framework around your model. It can use OpenAI Realtime, so the two can be part of the same build. | Similar product Google Cloud platform for deterministic and generative voice agents built with flows, playbooks, data stores, and integrations. |
|---|---|---|---|
| Starting price | From $32/M tokens | Usage-based | From $0.06/min |
| Free to start | Not published | Free plan | Free credits |
| Calls | Inbound & outbound | Inbound & outbound | Inbound |
| HIPAA | Enterprise plans | Enterprise plans | BAA available |
| Morak scorecard How we score | |||
| Ease of setup | 2 of 5, Developer build | 2 of 5, Developer build | 2 of 5, Developer build |
| Pricing transparency | 5 of 5, Plans published | 2 of 5, No public rates | 5 of 5, Rates published |
| Try before you buy | 3 of 5, Web demo | 5 of 5, Free plan | 4 of 5, Self-serve trial |
| Call & telephony control | 4 of 5, Broad | 5 of 5, Full control | 4 of 5, Broad |
| Compliance readiness | 5 of 5, Strong | 5 of 5, Strong | 5 of 5, Strong |
| Integration coverage | 3 of 5, 2 of 4 covered | 3 of 5, 2 of 4 covered | 4 of 5, 3 of 4 covered |