OpenAI has released two new speech-to-speech models that make real-time voice agents significantly faster and cheaper to run. On July 6, 2026, the company shipped gpt-realtime-2.1 and gpt-realtime-2.1-mini to its Realtime API, cutting response latency by at least 25% and reducing the cost of repeated input tokens by up to 80% through improved caching. For teams building customer support bots, voice assistants, and interactive applications, this update reshapes the economics of production voice systems.

What Changed in OpenAI's Latest Voice Models?

The two new models target the same job: powering low-latency voice agents that listen and speak in real time over the same Realtime endpoint developers already use. According to OpenAI's announcement, gpt-realtime-2.1 brings improved recognition of alphanumeric sequences like order numbers and confirmation codes, better handling of silence and background noise, and more graceful interruption behavior when a caller talks over the agent. The model also includes configurable reasoning effort, instruction following, and tool use capabilities.

gpt-realtime-2.1-mini is designed as a smaller reasoning model for faster, lower-cost interactions, offering the same capability envelope but tuned for volume and price. Both models are available now to test in OpenAI's Realtime playground.

How Does the Latency Improvement Actually Work?

In voice interfaces, latency is the product. A 25% reduction in p95 latency (the slowest one-in-twenty responses) is the difference between an assistant that feels like a phone tree and one that feels like a person. OpenAI says the gain comes from improved caching rather than a new model architecture, which reveals the real optimization strategy.

Voice agents resend a lot of the same context on every turn: the system prompt, tool definitions, and the growing transcript of the conversation. If those repeated tokens are served from cache instead of recomputed, both the time-to-first-token and the cost per turn fall. That is why the latency win and the pricing win in this release are really the same story told twice.

What Are the Pricing Changes, and Why Do They Matter?

The most consequential number in this release is not the latency figure. It is the cost of cached input tokens. On gpt-realtime-2.1, cached text input is billed at $0.40 per million tokens, one tenth of the $4.00 per million charged for uncached text input. Cached audio input gets the same $0.40 per million rate, against $32.00 per million for standard audio input, an 80-fold discount on the tokens a voice agent repeats most.

For a customer-support agent that streams audio in and out while resending a fat system prompt every turn, caching turns the dominant cost line into a rounding error, while output tokens become the real budget. That inverts how teams should think about prompt design for voice: long, static instructions are now cheap to keep resident, so the optimization moves to keeping responses tight.

How to Optimize Voice Agents for the New Pricing Model

• Leverage Cached Inputs: Store system prompts, tool definitions, and conversation context in cache to reduce costs by up to 80% on repeated tokens, making long static instructions economical to maintain.
• Tighten Output Responses: Since output tokens now carry the highest cost ($24 per million for text, $64 per million for audio on the flagship), focus optimization on keeping responses concise and direct.
• Choose the Right Reasoning Effort: Start with low reasoning effort for most production voice agents, then tune up or down based on task complexity, latency tolerance, and failure cost to balance speed and accuracy.
• Select the Appropriate Model Tier: Use gpt-realtime-2.1-mini for high-volume, lower-complexity voice work where cost matters most, and reserve the flagship for tasks requiring higher reasoning capability.

gpt-realtime-2.1-mini is OpenAI's answer to a market that has been undercutting it on price. Across the board it costs a fraction of the flagship: $10.00 per million audio input tokens versus $32.00, $20.00 per million audio output tokens versus $64.00, and $2.40 per million text output tokens versus $24.00, roughly a third of the audio cost and a tenth of the text-output cost of gpt-realtime-2.1. For high-volume, lower-complexity voice work, that is the tier most builders will reach for first.

What Is the Reasoning Effort Setting, and When Should You Use It?

gpt-realtime-2.1 exposes a configurable reasoning effort setting, the same knob OpenAI added to its reasoning models, now wired into speech-to-speech. The allowed values are minimal, low, medium, high, and xhigh, and the guidance is explicit: start at low for most production voice agents, then tune up or down based on task complexity, latency tolerance, and failure cost.

OpenAI's use-case ladder maps reasoning effort to real-world tasks. Minimal suits smart-home commands, timers, and simple lookups where speed is everything. Low covers customer support, order lookups, and basic policy questions. Medium is for multi-step reasoning like technical support and diagnostics. High is for high-precision or escalation decisions, and xhigh is reserved for complex planning and high-stakes orchestration where the added latency is worth it. For voice agents specifically, the default instinct should be to stay low and only climb when a task genuinely needs it, a discipline that matters more in a live conversation than in a chat window, where a pause is far more noticeable.

Where Does This Fit in the Broader Voice AI Market?

The context is a crowded, fast-moving field. Voice-agent platforms like Vapi build on top of models like these, while ElevenLabs pushes on voice quality and Google and Amazon push their own real-time speech stacks. OpenAI competes on two fronts at once: raw model quality against specialists like ElevenLabs and Google, and price against everyone. A cheaper mini tier is how it defends the high-volume middle of the market that actually pays the bills.

For developers and teams already using OpenAI's Realtime API, this update is a straightforward win worth adopting: lower latency means more natural conversations, and dramatically cheaper cached tokens reshape the budget for production voice systems. The combination of speed and cost efficiency positions these models as a competitive option for the growing wave of voice-first applications in customer service, healthcare, and beyond.