Published: July 7, 2026
On July 6, 2026, OpenAI released two new Realtime models on its API: gpt-realtime-2.1 and gpt-realtime-2.1-mini, both built for low-latency voice and multimodal applications. gpt-realtime-2.1 is the higher-capability model for realtime reasoning, tool use, and voice-agent workflows, while gpt-realtime-2.1-mini is a smaller reasoning model aimed at faster and lower-cost voice interactions. OpenAI states that the update reduces p95 latency by at least 25 percent across its Realtime voice models through improved caching.
What did OpenAI release with gpt-realtime-2.1 and gpt-realtime-2.1-mini?
OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6, 2026 as two new speech-to-speech models in the Realtime API. Both are designed for developers building voice agents and other low-latency voice and multimodal experiences. gpt-realtime-2.1 is positioned as the stronger of the two, with support for configurable reasoning effort, instruction following, and tool use, so a voice agent can call functions and follow multi-step instructions during a live conversation.
Gpt-realtime-2.1-mini is described by OpenAI as a mini reasoning model that trades some capability for lower cost and faster responses. It targets high-volume or latency-sensitive voice applications where speed and price matter more than maximum reasoning. Both models are available through the Realtime API and can be tested in the OpenAI Playground.
What is new in gpt-realtime-2.1 compared to GPT-Realtime-2?
Gpt-realtime-2.1 is an update to GPT-Realtime-2, the Realtime voice model OpenAI introduced in June 2026. According to OpenAI, gpt-realtime-2.1 improves alphanumeric recognition, which matters for reading back items such as order numbers, phone numbers, and confirmation codes, together with better silence and noise handling and more reliable interruption behavior when a user speaks over the model.
The release also lowers p95 latency by at least 25 percent across OpenAI’s Realtime voice models, an improvement OpenAI attributes to better caching rather than a change in model size. gpt-realtime-2.1 keeps the speech-to-speech design of its predecessor and adds configurable reasoning effort, so developers can set how much the model reasons before it responds. For teams already using GPT-Realtime-2, moving to gpt-realtime-2.1 is a version update within the same Realtime API rather than a new integration.
Gpt-realtime-2.1 and gpt-realtime-2.1-mini pricing
OpenAI prices both models per 1 million tokens, split across text, audio, and image inputs and outputs. The table below lists the published rates in US dollars.
| Price (per 1M tokens, USD) | Gpt-realtime-2.1 | Gpt-realtime-2.1-mini |
|---|---|---|
| Text input | $4.00 | $0.60 |
| Text cached input | $0.40 | $0.06 |
| Text output | $24.00 | $2.40 |
| Audio input | $32.00 | $10.00 |
| Audio cached input | $0.40 | $0.30 |
| Audio output | $64.00 | $20.00 |
| Image input | $5.00 | $0.80 |
| Image cached input | $0.50 | $0.08 |
The largest difference is in audio. gpt-realtime-2.1-mini charges $10.00 for audio input against $32.00 for gpt-realtime-2.1, and $20.00 for audio output against $64.00. That gap is the main reason to consider the mini model for high-volume voice traffic.
How to access the models and which one to choose
Both models are available now through OpenAI’s Realtime API, using the model IDs gpt-realtime-2.1 and gpt-realtime-2.1-mini, and can be tried in the Realtime section of the OpenAI Playground. OpenAI’s guidance is direct: use gpt-realtime-2.1 when the application needs the strongest realtime reasoning, tool use, instruction following, and voice-agent behavior, and use gpt-realtime-2.1-mini when speed and cost matter more than maximum capability.
Full model details and pricing are in OpenAI’s announcement: https://community.openai.com/t/new-realtime-models-on-the-api-gpt-realtime-2-1-and-gpt-realtime-2-1-mini/1385896
