OpenAI Launches GPT Realtime 2 for Voice Agents
OpenAI released GPT Realtime 2 on May 7, 2026. This update includes configurable reasoning for speech-to-speech agents and new streaming translation models.
Release Details and Scope
OpenAI officially released GPT Realtime 2 on May 7, 2026. This date marks the historical announcement of the feature set described in the official changelog. The release introduces gpt-realtime-2 as a new realtime voice model designed for speech-to-speech agents. A key characteristic noted in the source is that this model supports configurable reasoning. Alongside the primary model, OpenAI released two specialized variants: GPT Realtime Translate for streaming speech translation and GPT Realtime Whisper for streaming speech-to-text transcription.
- Primary model: gpt-realtime-2 with configurable reasoning.
- Specialized model: gpt-realtime-translate for streaming translation.
- Specialized model: gpt-realtime-whisper for streaming transcription.
- Input modalities include text and audio; output modalities include text and audio.
API Endpoints and Documentation Updates
The release accompanies specific API updates to support these new capabilities. The v1/realtime endpoint is associated with the main model, while v1/realtime/translations supports the translation variant. Additionally, v1/realtime/transcriptionsessions handles the transcription workflow. Documentation has been updated to reflect these changes. OpenAI refreshed the Realtime and audio guide, added a dedicated guide for Realtime translation, and updated the Realtime transcription section for streaming transcripts. The source notes that realtime prompting guidance was moved within the documentation structure. These updates provide engineers with structured resources for integrating the new models into existing systems.
Evaluation Procedure and Limitations
Engineers can evaluate the new models by designing controlled tests against their specific use cases. To validate performance, one should test the gpt-realtime-2 model with varying levels of configurable reasoning to measure latency and accuracy in speech-to-speech interactions. For the translation component, measure the fidelity of streaming translations using gpt-realtime-translate against known reference pairs. Similarly, validate the transcription accuracy of gpt-realtime-whisper by comparing its output against ground-truth text from diverse audio samples. Compare these metrics against baseline requirements or previous internal benchmarks if available. It is important to note that the source does not confirm open weights, pricing details, context window limits, or maximum output tokens. Therefore, any assumptions about local installation, cost structures, or architectural specifics remain unverified. The evaluation must rely solely on the documented API behavior and observed outputs during testing.
Primary sources
Frequently asked questions
What specific features are included in the GPT Realtime 2 release?
The release includes gpt-realtime-2 with configurable reasoning for speech-to-speech agents, gpt-realtime-translate for streaming speech translation, and gpt-realtime-whisper for streaming speech-to-text. It also includes updated API endpoints and documentation guides for Realtime translation and transcription.
Does the source confirm whether GPT Realtime 2 has open weights?
No. The source records that open weights are not confirmed (openWeights=false). This does not establish that local installation is impossible or that API access is mandatory; it simply indicates that open weight availability is not verified in the provided data.