OpenAI Realtime

OpenAIRealtime is the engine wrapper for OpenAI’s Realtime API — a single WebSocket session that handles speech-in, reasoning, and speech-out, with sub-500 ms typical turn latency. For the basic engine=OpenAIRealtime(...) quickstart, see Engines. This page documents the full configuration surface: every supported model, the streaming transcription options, and the new reasoning_effort tier.

Models

Pass any of these to model= on OpenAIRealtime(...). Pricing is auto-resolved per model from DEFAULT_PRICING — no manual override is required (see Metrics).

Model	Audio in / out (per M tokens)	Notes
`"gpt-realtime-mini"` (default)	$10 /$ 20	Fastest + cheapest. Production default for most voice flows.
`"gpt-realtime"`	$32 /$ 64	GA realtime model (Aug 2025).
`"gpt-realtime-2"`	$32 /$ 64	Most-capable. Stronger instruction following, 128K context, supports `reasoning_effort`.
`"gpt-4o-realtime-preview"`	$100 /$ 200	Earlier preview, retained for compatibility.
`"gpt-4o-mini-realtime-preview"`	$10 /$ 20	Earlier preview, retained for compatibility.

The same identifiers are exposed as a StrEnum for editor autocomplete:

from getpatter.providers.openai_realtime import OpenAIRealtimeModel

OpenAIRealtimeModel.GPT_REALTIME_2  # "gpt-realtime-2"

gpt-realtime-translate is intentionally not supported by Patter’s Realtime engine. It lives on a different OpenAI endpoint (/v1/realtime/translations), does not accept tool calls or response.create, and would invalidate the Agent contract Patter exposes. Real-time translation, if added, will land as a dedicated feature — not as a Realtime model variant.

Reasoning effort

gpt-realtime-2 accepts a configurable reasoning tier. Patter exposes it as the reasoning_effort constructor argument on the lower-level OpenAIRealtimeAdapter:

Value	When to use
`"minimal"`	Snappy turn-taking. Skips most reasoning.
`"low"`	Recommended for production voice. Good instruction following without measurable per-turn latency.
`"medium"`	Multi-step tool flows where the model should plan. Adds latency.
`"high"`	Complex reasoning. Not recommended for live phone calls.

When set, Patter injects session.reasoning = { effort: ... } into the session.update payload. When omitted, the field is not sent and OpenAI’s server default applies. The field is a no-op on models that ignore it (for example gpt-realtime-mini), so it’s safe to leave configured across model swaps.

Higher reasoning tiers add measurable latency to every turn. Stick to "low" unless you’ve profiled the call and confirmed the model needs more.

Streaming transcription

The Realtime session can run an inline Whisper-family model on inbound audio so you get text deltas alongside the conversation. The model is set via input_audio_transcription_model:

Model	Cost	Notes
`"whisper-1"` (default)	$0.006/min	Established Whisper. Slower partials.
`"gpt-4o-mini-transcribe"`	$0.003/min	Cheapest.
`"gpt-4o-transcribe"`	$0.006/min	Higher accuracy.
`"gpt-realtime-whisper"`	$0.017/min	Streaming-optimised. Lowest-latency partials. Use when you need fast deltas in the dashboard or for live captioning.

Same enum form:

from getpatter.providers.openai_realtime import OpenAITranscriptionModel

OpenAITranscriptionModel.GPT_REALTIME_WHISPER  # "gpt-realtime-whisper"

Speakerphone noise & turn detection

The same noise_reduction and turn_detection knobs documented on the OpenAI Realtime 2 page are accepted by OpenAIRealtime(...) too — they reach the beta adapter, which emits input_audio_noise_reduction at the top level of the session (the GA adapter nests it under session.audio.input):

from getpatter import OpenAIRealtime, RealtimeTurnDetection

engine = OpenAIRealtime(
    noise_reduction="far_field",  # filter speakerphone / room noise
    turn_detection=RealtimeTurnDetection(type="server_vad", threshold=0.6),
)

None / omitting either field preserves today’s behaviour exactly. See the Realtime 2 reference for the full RealtimeTurnDetection field table.

Worked example — `gpt-realtime-2` with low reasoning + streaming whisper

Constructing the lower-level OpenAIRealtimeAdapter directly gives access to every field. This is what OpenAIRealtime(engine=...) builds under the hood; reach for it when you need reasoning_effort or a non-default transcription model.

import asyncio

from getpatter import Patter, Twilio
from getpatter.providers.openai_realtime import (
    OpenAIRealtimeAdapter,
    OpenAIRealtimeModel,
    OpenAITranscriptionModel,
)

phone = Patter(carrier=Twilio(), phone_number="+15555550100")  # TWILIO_* from env

adapter = OpenAIRealtimeAdapter(
    api_key="",                                            # reads OPENAI_API_KEY
    model=OpenAIRealtimeModel.GPT_REALTIME_2,
    voice="nova",
    instructions="You are a helpful, concise voice assistant.",
    input_audio_transcription_model=OpenAITranscriptionModel.GPT_REALTIME_WHISPER,
    reasoning_effort="low",                                # OpenAI's recommended production tier
)

agent = phone.agent(
    engine=adapter,                                        # adapter passed as engine
    system_prompt="You are a helpful assistant.",
    first_message="Hi, how can I help today?",
)

async def main() -> None:
    await phone.serve(agent)

asyncio.run(main())

// See /typescript-sdk/providers/openai-realtime

Since 0.6.2 you can pass reasoning_effort and input_audio_transcription_model directly to the engine wrapper — OpenAIRealtime(model="gpt-realtime-2", reasoning_effort="low", input_audio_transcription_model="gpt-realtime-whisper"). Reach for the lower-level OpenAIRealtimeAdapter only when you need every field (custom VAD type, modalities, silence_duration_ms, etc.). For the GA gpt-realtime-2 endpoint, prefer the dedicated OpenAIRealtime2 marker — it dispatches to a separate adapter that handles the GA-shape session.update wire format automatically.

Backward compatibility

Defaults are unchanged: model="gpt-realtime-mini", input_audio_transcription_model="whisper-1", reasoning_effort=None.
All existing OpenAIRealtime(...) constructions keep working without code changes.
Pricing for new models is added under DEFAULT_PRICING["openai_realtime"].models[...]. The earlier Patter(pricing={"openai_realtime": DEFAULT_PRICING["openai_realtime_2"]}) workaround is no longer needed — just construct with model="gpt-realtime-2".

What’s Next

Engines

All engine classes side by side.

Metrics

Per-call cost breakdown and the model-aware pricing table.

Agents

Configure system prompts, tools, and first messages.

Tools

Function calling inside a Realtime session.

Get Started

Setting up Patter

Observability

Integrations

Development

OpenAI Realtime

OpenAI Realtime

Models

Reasoning effort

Streaming transcription

Speakerphone noise & turn detection

Worked example — `gpt-realtime-2` with low reasoning + streaming whisper

Backward compatibility

What’s Next

Engines

Metrics

Agents

Tools

​OpenAI Realtime

​Models

​Reasoning effort

​Streaming transcription

​Speakerphone noise & turn detection

​Worked example — gpt-realtime-2 with low reasoning + streaming whisper

​Backward compatibility

​What’s Next

Engines

Metrics

Agents

Tools

OpenAI Realtime

Models

Reasoning effort

Streaming transcription

Speakerphone noise & turn detection

Worked example — `gpt-realtime-2` with low reasoning + streaming whisper

Backward compatibility

What’s Next