Vendor intelligence
OpenAI Realtime API
Speech-to-speech API for developers implementing their own voice application.
Read the signals. Keep the scope.
A place in your stack.
Best for
- Custom agents that need direct realtime model access and tool orchestration.
- Teams able to instrument token usage and implement carrier and business-system logic.
Watch-outs
- A token rate cannot be compared directly with a bundled call-minute price.
- Pin the model and track conversation context, caching and transcription costs before estimating a production bill.
Product specification
What the evidence says.
Unknown means we have not established it. It does not mean No. Compliance claims require plan and contract review.
White label
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)Client tenancy
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)Client billing
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)ISO 27001
Not established by the evidence reviewed. This does not mean unsupported.
No verified sourceBAA availability
General healthcare-addendum guidance does not alone establish the exact Realtime model/endpoint/region agreement for this buyer; confirm covered service and contract.
Source ↗ (opens in new tab)SIP
Realtime section documents incoming SIP calls through a buyer-selected provider.
Source ↗ (opens in new tab)Warm transfer
SIP REFER is established, but consultation, context payload, queue routing, summary and transfer-failure recovery are not specified as Realtime features.
Source ↗ (opens in new tab)Cold transfer
Realtime call refer endpoint sends SIP REFER to target; successful API response is not transfer completion proof.
Source ↗ (opens in new tab)Batch calling
Not established by the evidence reviewed. This does not mean unsupported.
No verified sourceMCP tools
Remote MCP tools and application functions; connector_id deprecated for models released after September1,2026.
Source ↗ (opens in new tab)Webhooks
Incoming-call webhook with signature verification and call lifecycle API.
Source ↗ (opens in new tab)Calendar
Legacy built-in read connector on eligible pre-cutoff model; not native booking or current-model blanket support.
Source ↗ (opens in new tab)CRM
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)Bring your carrier
Realtime section documents incoming SIP calls through a buyer-selected provider.
Source ↗ (opens in new tab)Bring your model
Native audio model is documented. Custom tools/backend code do not prove a supported replacement for the native TTS/STT/LLM stages.
Source ↗ (opens in new tab)Bring your voice
Native audio model is documented. Custom tools/backend code do not prove a supported replacement for the native TTS/STT/LLM stages.
Source ↗ (opens in new tab)HIPAA / BAA
General healthcare-addendum guidance does not alone establish the exact Realtime model/endpoint/region agreement for this buyer; confirm covered service and contract.
Source ↗ (opens in new tab)SOC 2
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)GDPR
No supporting statement located in the bounded official corpus; confirm scope with vendor. Unknown is not No.
Source ↗ (opens in new tab)Languages
Transcription language hints and examples do not establish a complete Realtime conversational language matrix. Translation model language counts cannot be reused.
Source ↗ (opens in new tab)Source ↗ (opens in new tab)Voices
Not established by the evidence reviewed. This does not mean unsupported.
No verified sourceRealtime audio
No independent latency or quality measurement.
Source ↗ (opens in new tab)Voice cloning
Restricted availability; consent and sample recordings required.
Source ↗ (opens in new tab)Architecture & connections
Understand the moving parts.
STT
OpenAI configured input transcription models
LLM
OpenAI
TTS
OpenAI native Realtime audio
TELEPHONY
External SIP carrier
Integrations
Function tools · Remote MCP · Webhooks
Evidence you can inspect.
A documented feature, a supplier claim and an unanswered question are different kinds of evidence.
Evidence coverage details
- Strong-source resolved
- 31
- Vendor-claim only
- 0
- Conditional
- 17
- Other sourced outcomes
- 0
- Unknown after research
- 66
- Conflicting evidence
- 0
- Not applicable
- 0
- Not researched
- 0
Inspect detailed evidence, sources and research gaps
September 2026 remains the historical baseline. A new research date does not mean a feature launched that day. Unknown after research and not researched are separate states.
Official documentation establishes what the supplier documents. No paid calls, independent product benchmarks or compliance audits were performed in this research pass. Coverage measures research work, never product quality.
The operating decision.
Price structure
Component stack
V3.3 evidence review · GPT-Realtime-2 input audio tokens only; output/audio/text/cache have different rates; not 2.1 rate evidence
Realtime model tokens plus input transcription, tools and carrier/hosting; not GPT-Live session-minute billing.
Human fallback
Unknown — no established fallback mode
No individually verified capability establishes this selection. Conditional support and vendor claims do not satisfy Strict Must-have.
Documented transfer is not a staffed human service. Verify who answers, when they answer and what happens if they do not.
Try it yourself
Unknown — no independently checked public demo entry is attached.
External signals
No dated external signal in this record. Missing evidence is not a negative review.
Underlying model context: the documented stack above does not establish an exact benchmark checkpoint or its deployed settings. Public model results cannot become a OpenAI Realtime API score.
Append-only price history
Every capture leaves a trace.
A new check creates a new snapshot. Deltas appear only for the same currency, unit, model and scope.
Tracked: $32 / 1m-tokens · GPT-Realtime-2 audio input only; output and other modalities are additional
Tracked: $32 / 1m-tokens · V3.3 evidence review · GPT-Realtime-2 input audio tokens only; output/audio/text/cache have different rates; not 2.1 rate evidence
Tracked: $32 / 1m-tokens · V3.3 evidence review · GPT-Realtime-2 input audio tokens only; output/audio/text/cache have different rates; not 2.1 rate evidence
New research observations preserve earlier captures. “Evidence improved” records what we learned; it does not claim the market changed on that date. September 2026 is the historical baseline.
Price the actual workflow.
Complete cost cannot be established from this reference. Full bill requires measured model/token/cache/tool/carrier usage; conversation minutes alone are insufficient. · Input transcription, tool/backend calls, carrier and application infrastructure
Carry the current source into your estimate, then enter the model, voice, carrier and fixed fees your configuration needs.
Alternatives by category
Other developer frameworks to inspect.
Alphabetical alternatives within the same category. An editorial starting point, not a performance ranking.
Teams that want to own pipeline code and choose deployment infrastructure.
CompareEvidence archive
Inspect the record.
GPT-Realtime-2 model pricingreviewed
GPT-Realtime-2: per million audio tokens, input $32, cached input $0.40 and output $64; text input $4 and output $24. The free API usage tier is unsupported. These are token rates, not an all-in call-minute price.
Inspect official source ↗OpenAI Realtime telephony and SIPreviewed
The Realtime API accepts SIP calls using incoming-call webhooks. The call refer endpoint forwards SIP REFER to the carrier; the downstream system completes the transfer. Carrier setup remains external.
Inspect official source ↗OpenAI Realtime tools and MCPreviewed
Realtime sessions can call function tools and remote MCP servers. Remote tools require configuration, authentication where relevant, and application handling of lifecycle events.
Inspect official source ↗OpenAI custom voicesreviewed
Custom voices are limited to eligible customers. Access requires contacting sales; voice creation requires a consent recording and a voice sample.
Inspect official source ↗OpenAI voice cost guidancereviewed
Conversational Realtime charges for input and output tokens across modalities. Input transcription uses a separate model and rate card. Conversation growth, caching and reasoning settings affect cost.
Inspect official source ↗Evidence describes documented scope, not a tested platform score. Unknown means a conclusion has not been established; API connectivity does not establish a native integration. Read our research methodology →
Public affiliate terms and an approved tracking link have not been verified.
Partner eligibility and commercial terms are unverified.