Voice agents have always had a tell. You ask one to do something real — pull up an order, check an invoice — and it goes silent mid-task. You assume the line dropped. You start talking again. Now the model is confused, the tool call is half-finished, and the conversation is a mess. On July 6, 2026, OpenAI shipped the fix for that specific failure, and it did it in the cheap tier.
GPT-Realtime-2.1 and GPT-Realtime-2.1-mini are the new Realtime API models, and the headline is not raw speed — it's reasoning. The mini model can now think before it speaks, call your functions, and narrate what it's doing while it works. That last part sounds small. It is the difference between a voice agent that feels broken and one that feels competent.
What actually changed
The Realtime API runs audio through a single model rather than chaining a speech-to-text engine, an LLM, and a text-to-speech engine together. That design already cut latency and preserved tone. The 2.1 release adds two things on top of it.
First, reasoning reaches the mini tier. gpt-realtime-2.1-mini is a mini reasoning model — it can plan internally, call a tool, and then answer, all over a live audio connection. Reasoning effort is configurable across five levels: minimal, low, medium, high, and xhigh. Low is the default, and OpenAI explicitly advises starting there for most production voice agents. Higher effort buys you depth on hard turns at the cost of latency and output tokens.
Second, the larger gpt-realtime-2.1 sharpens the fundamentals: better alphanumeric recognition, improved silence and noise handling, and more reliable interruption behavior when a caller talks over the model.
Why alphanumeric recognition matters: reading back an order number like "8-3-5-7-1" and getting it right on the first pass is the entire job for a field-service or support agent. A single transposed digit routes the whole call into failure.
The latency and caching story
The performance claim is a cut of at least 25% in p95 latency across the Realtime voice models. That's the 95th-percentile response time — the slow tail users actually feel on a live call, not the median that looks good in a demo. OpenAI says the improvement came from better caching at the serving layer, not from changing model size.
Caching does double duty here. It lowers the slow-tail latency and it lowers cost, because cached input tokens are billed at a steep discount. For the mini model, cached audio input drops to $0.30 per 1M tokens, versus $10.00 per 1M for fresh audio input. Long sessions benefit most: your system prompt caches after the first turn and rides free for the rest of the call.
The pricing, plainly
Pricing is per 1M tokens, split across text, audio, and image. Here's where the two models sit:
| Price (per 1M tokens) | gpt-realtime-2.1 |
gpt-realtime-2.1-mini |
|---|---|---|
| Reasoning | Yes (configurable effort) | Yes (mini reasoning model) |
| Audio input | $32.00 | $10.00 |
| Audio input (cached) | $0.40 | $0.30 |
| Audio output | $64.00 | $20.00 |
| Image input | $5.00 | $0.80 |
The clever part: the mini keeps the exact rate of the previous gpt-realtime-mini while adding reasoning on top. You get a genuinely new capability at the old price. On audio output — usually the dominant cost in a voice call — the full model runs roughly 3x the mini's rate. For most support and assistant workloads, the mini at low effort is the right default, and you reach for the full model only when a task genuinely needs stronger reasoning or instruction following.
Wiring it up
The connection model depends on where your audio lives. Browser clients connect over WebRTC, with your server minting a short-lived client secret first so your real API key never touches the client. Server-side media pipelines use WebSockets, and telephony integrations use SIP. A typical mini session config sets the model, reasoning: { effort: "low" }, and your tool definitions, then hands the browser an ephemeral secret to connect directly.
OpenAI's own migration advice is sensible: start reasoning effort at low, raise it only for harder tasks, separate hard rules from soft defaults in your instructions, and run evals before and after any model swap. Audio-token pricing is genuinely hard to convert into a per-call number, so measure real sessions rather than trusting a back-of-envelope estimate.
The Bottom Line
GPT-Realtime-2.1 is not a flashy launch, and that's the point. Adding configurable reasoning to a low-cost voice tier, cutting p95 latency 25% through caching, and holding the mini price flat is exactly the kind of unglamorous infrastructure work that makes production voice agents viable instead of demo-only. If you shipped anything on the older Realtime models, this is a low-risk upgrade with a real payoff — just run your evals first and start at low effort.
