OpenAI’s new voice AI models slash latency and costs for real-time agents
OpenAI has just pushed two fresh voices into its Realtime API lineup, promising crisper, faster, and cheaper conversational agents. The new GPT-Realtime-2.1 and its smaller sibling GPT-Realtime-2.1-mini arrive with a sharper ear for alphanumeric cues and a knack for staying audible while they work. Both models also deliver at least a quarter less p95 latency, thanks to tighter caching that cuts both wait times and expenses for long-running sessions.
A single brain, not a pipeline
The Realtime API sidesteps the usual speech-to-text then text-to-speech detour by running everything through one model, keeping turns snappy and tonal nuances intact. GPT-Realtime-2.1-mini brings reasoning under the same hood, letting it plan a step, call an external tool, then speak its answer without ever falling silent mid-task. That matters because voice agents used to freeze during tool calls, making users wonder if the line dropped. The mini model answers with a spoken preamble—“I’ll check that order now”—and keeps talking while it works, keeping multi-step flows coherent.
Tunable effort, clearer bills
Developers can dial reasoning effort from minimal to extra-high, balancing speed against depth. Low effort keeps latency low for routine turns, while higher levels burn more tokens and time. OpenAI suggests starting low for most production bots. On the pricing side, the real win is smarter caching. Cached audio input for the mini version drops to $0.30 per 1M tokens, versus $10.00 for fresh audio, and cached text tokens cost $0.06, a tenth of live input. Sessions benefit after the first turn, when the system prompt locks into cache.
Which model for which job?
GPT-Realtime-2.1 steps up where you need the strongest real-time reasoning, tool use, instruction following, and voice-agent behavior. Its mini counterpart offers the same capabilities at lower cost and latency, making it a natural fit for high-volume, low-friction voice assistants. Together, they give developers more knobs to tune performance and pricing without rewriting the stack.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.