OpenAI Ships GPT-Realtime-2.1: Reasoning Comes to Voice Agents, p95 Latency Drops 25%
OpenAI published two new Realtime API models on July 6 targeting low-latency voice and multimodal agent builds: gpt-realtime-2.1 and gpt-realtime-2.1-mini. Both models ship a 25% reduction in p95 latency across realtime voice interactions, credited to improved caching architecture. The more significant change is the mini model: it is the first mini reasoning model designed specifically for real-time voice, adding chain-of-thought capability to voice agents without raising the price.
What Changed in Each Model
gpt-realtime-2.1 updates the previous gpt-realtime-2 with three targeted improvements: better alphanumeric recognition (phone numbers, addresses, codes read aloud), more reliable silence and noise handling, and cleaner interruption behaviour when users cut off mid-sentence. Critically, it adds configurable reasoning effort — developers can dial compute up or down on the reasoning chain before the model responds. That is new territory for a realtime model.
gpt-realtime-2.1-mini ships at the same cost as gpt-realtime-mini, meaning the mini reasoning capability arrives at no premium. It uses a mini reasoning model rather than configurable-effort reasoning, making it suited for lower-latency applications where the reasoning task is simpler.
Pricing
| gpt-realtime-2.1 | gpt-realtime-2.1-mini | gpt-realtime-mini (prior) | |
|---|---|---|---|
| Text input | $4.00/M | $0.60/M | $0.60/M |
| Text cached input | $0.40/M | $0.06/M | $0.06/M |
| Text output | $24.00/M | $2.40/M | $2.40/M |
| Audio input | $32.00/M | $10.00/M | $10.00/M |
| Audio cached input | $0.40/M | $0.30/M | $0.30/M |
| Audio output | $64.00/M | $20.00/M | $20.00/M |
| Image input | $5.00/M | $0.80/M | $0.80/M |
The 3x audio output gap between the full model ($64/M) and the mini ($20/M) is the primary decision lever. Teams can accept reduced reasoning depth in exchange for a 70% cut to audio output costs.
Why Reasoning in Voice Matters
The previous generation of realtime voice models were reactive. They transcribed, generated, and returned audio — the intelligence was fast, but not deliberative. Adding configurable reasoning effort to gpt-realtime-2.1 means the model can now pause internally to reason through a multi-step question before committing to an audio response. For voice agents handling complex queries — diagnostics, instructions, troubleshooting — that difference between “answer fast” and “answer correctly” is what most production deployments have been waiting for.
The latency cut matters independently. A 25% reduction in p95 latency from caching means tail-end slowdowns — the worst-case delays users notice in conversation — shrink. The improvement comes before any model-quality factor, simply from how frequently the system can serve from a cached state rather than a cold computation.
Context
This release follows the Artificial Analysis Speech-to-Speech Index, where GPT-Realtime-2 led at 77.2% and Grok Voice topped two of three sub-dimensions. The new models haven’t yet appeared on that leaderboard. The mini reasoning capability also arrives weeks after xAI shipped Grok Voice Think Fast 1.0 — which posted 70% autonomous resolution at Starlink — signaling OpenAI is responding to competitors adding cognitive depth to voice at the infrastructure level.