OpenAI Ships GPT-Realtime-2.1 with 25% Lower Latency and a Reasoning Mini Tier

OpenAI released two new voice models for its Realtime API on July 6: gpt-realtime-2.1 and a smaller gpt-realtime-2.1-mini that gains reasoning capability for the first time, at the same price as its predecessor.

• P95 latency across Realtime voice models drops by at least 25% via improved caching
• gpt-realtime-2.1: $4.00 / $32.00 / $64.00 per million tokens for text input / audio input / audio output
• gpt-realtime-2.1-mini: $0.60 / $10.00 / $20.00 per million tokens, unchanged from the prior mini
• Mini tier adds configurable reasoning effort for realtime voice agents
• Full model improves alphanumeric recognition, noise handling, and interruption behavior

The update targets production voice agents: both models can narrate intermediate steps, such as saying 'I'll check that order now,' before executing a tool call, to keep multi-step conversations coherent. Compare voice and reasoning models side by side in the arena.