Access to this page requires authorization. You can try signing in or changing directories . Access to this page requires authorization. You can try changing directories . Currently viewing: Foundry (classic) portal version - Switch to version for the new Foundry portal The generally available GPT Realtime 2.1 models are speech-to-speech models with built-in reasoning. They accept audio input and produce audio output. They're designed for low-latency, interactive voice experiences where you need stronger instruction following and reasoning than earlier realtime models. The current models are: gpt-realtime-2.1 (version 2026-07-07 ) is an incremental update to gpt-realtime-2 with improved silence and noise handling. gpt-realtime-2.1-mini (version 2026-07-07 ) is a smaller variant of GPT Realtime 2.1. Both models support Global Standard and Data Zone Standard deployments. For region-by-region availability, see Region availability for Foundry Models sold by Azure . Azure OpenAI bills these models by input, cached input, and output tokens. Image input is billed separately. For current rates, see the Audio Models section on the Azure OpenAI pricing page . GPT Realtime 2.1 features Reasoning support with an adjustable reasoning.effort control. Response phases that distinguish preambles ("commentary") from the final answer ("final_answer"). Longer context window (256,000 tokens). API and transports Use the GA Realtime API under /openai/v1 . The WebSocket endpoint has the following format, where the model value is your deployment name: The GA endpoint doesn't use a date-based API version or an api-version query parameter. For complete examples, see Use the GPT Realtime API via WebSockets and Use the GPT Realtime API via WebRTC . Audio and voices For audio sent as Realtime API events, use headerless, mono PCM16 audio sampled at 24 kHz and base64-encode the bytes. WebRTC negotiates audio transport between the client and service. The supported output voices are alloy , ash , ballad , coral , cedar , echo , marin , sage , shimmer , and verse .

Reasoning effort Control reasoning intensity with the reasoning.effort session parameter. Valid values are minimal , low , medium , and high . Preambles and response phases Realtime responses can include multiple output items per turn. Each item has a phase that indicates its role: Preambles can reduce perceived latency. For example, the model might say, "Let me think about that..." They can also be used for tool announcements or silence fillers. If the model is interrupted during thinking, it discards the current chain of thought and starts a new turn. Instruction following Instruction following is stricter than in earlier realtime models. If your system prompt contains narrow wording (for example, distinguishing "order ID" from "confirmation code"), you might need to broaden or rephrase instructions to match real user phrasing. Known limitations Currently, GPT Realtime 2.x models don't support the truncation property in the session.update payload. To manage token usage and costs in an ongoing conversation, consider using the conversation.item.truncate or conversation.item.delete events. Realtime models have separate audio-token and concurrent-session quotas. Review Azure OpenAI quotas and limits before production deployment. The connection and usage patterns for GPT Realtime 2.1 are the same as for earlier versions. Deploy the new model and point your existing code to it. Choose the transport that fits your scenario: Use the GPT Realtime API via WebSockets Use the GPT Realtime API via WebRTC Use the GPT Realtime API via SIP Was this page helpful? Need help with this topic? Want to try using Ask Learn to clarify or guide you through this topic? Additional resources Last updated on 2026-10-23