GPT-Realtime-2.1 Playground: How to Test Voice Agents Step by Step

THE SHORT ANSWER

How do you use the GPT-Realtime-2.1 Playground?

START HERE · DIRECT SEARCH INTENT

GPT-Realtime-2.1 Playground: exact setup steps

1. Open the official entry point. Visit the GPT-Realtime-2.1 model page and use its “Try in Playground” action. We link to the documented entry point rather than guessing an undocumented deep link.
2. Confirm access. Sign in to your OpenAI API account, check model availability, budget and limits. The model card does not list Free-tier API support.
3. Choose the realtime voice experience. If the interface offers model selection, verify the exact model ID is gpt-realtime-2.1 rather than an older Realtime model or gpt-live-1.
4. Grant microphone permission. Use headphones where possible to reduce echo; test in a quiet room before introducing background noise.
5. Use one constrained instruction. Start with the SXF starter prompt below. Verify what the system says when information is missing instead of rewarding confident guesses.
6. Run a 60–90 second task. Include a serial number, a correction, and one hypothetical booking action that must not happen without confirmation.
7. Save evidence. Record scenario ID, model ID, prompt version, device, environment, output and any tool state. Export usage from your available telemetry rather than estimating tokens from call duration.

Playground controls and layout may change. The exact control labels after sign-in are account/interface dependent; the official model card is the source for the entry point.

COPY-READY · EXPERIMENT 01

First voice-agent test: a prompt that exposes real weaknesses

A good first test is not “Hello, how are you?” It should challenge entity recognition, clarification, interruption and authorization in the same short session. Use an invented customer and no production systems.

You are a careful voice assistant in a controlled evaluation.
Start with a short greeting. Ask one question at a time.
Before using any tool, restate the intended action and ask for confirmation.
Never claim a booking, refund, email, or database change happened unless a tool response confirms it.
If a name, date, code, or price is unclear, ask for clarification.
When interrupted, stop and respond to the user's latest correction.
Keep spoken answers concise, and explicitly state when you cannot verify a fact.

This prompt is an original SXF test design, not a claim that we executed a paid session or measured model performance. Run it repeatedly before drawing conclusions.

VERIFIED MODEL SPECIFICATIONS

What is GPT-Realtime-2.1—and what can it actually do?

GPT-Realtime-2.1 is an OpenAI reasoning-capable voice model for speech-to-speech interactions with tool use. The official model card attributes the update over Realtime-2 to improved alphanumeric recognition, silence/noise handling and interruption behavior. Those are vendor-reported changes, not independent SXF benchmark findings.

Read the authoritative limits on the OpenAI model card. For SXF’s broader model reference and release history, use our GPT-Realtime-2.1 model page.

IMPLEMENTATION · PRODUCTION ARCHITECTURE

How to connect GPT-Realtime-2.1 with WebRTC or WebSocket

The easiest Playground experiment requires no custom web app. Production architecture does. OpenAI’s current Realtime guidance uses a browser-facing WebRTC session with a short-lived client secret issued by your backend; server-managed audio can use WebSocket. Neither path requires putting your standard API key in public JavaScript.

Browser integration outline with the official Agents SDK

This is an integration outline based on the official Realtime API guide, not a copy-paste hosted service: the /api/realtime-token endpoint below is something you must implement on a trusted backend.

import { RealtimeAgent, RealtimeSession } from "@openai/agents/realtime";

// Your backend returns a short-lived realtime client secret.
// NEVER expose a standard OpenAI API key in browser JavaScript.
const res = await fetch("/api/realtime-token", { credentials: "same-origin" });
if (!res.ok) throw new Error("Unable to obtain ephemeral session key");
const { client_secret } = await res.json();
if (!client_secret) throw new Error("Missing ephemeral session key");

const agent = new RealtimeAgent({
  name: "SXF Voice QA Assistant",
  instructions: "Ask for confirmation before tool side effects; do not invent results.",
});

const session = new RealtimeSession(agent, {
  model: "gpt-realtime-2.1",
});
await session.connect({ apiKey: client_secret });

// Implement /api/realtime-token on a trusted server.
// Create the secret with POST /v1/realtime/client_secrets.
// This sample is an integration outline, not a hosted SXF voice demo.

The backend should authenticate the caller, enforce rate limits and issue short-lived credentials using the documented POST /v1/realtime/client_secrets. Never place OPENAI_API_KEY in static site assets, a client-side environment file, or a Playground prompt. See the current Realtime getting-started reference for supported session shape and the latest SDK.

TURN TAKING · INTERRUPTION EVIDENCE

Server VAD vs semantic VAD: what changes when someone interrupts?

Realtime VAD (voice activity detection) marks the beginning and end of user speech. OpenAI documents two supported turn-detection strategies for applicable Realtime sessions: server_vad detects pauses, while semantic_vad considers whether the speaker appears to have finished their thought. Neither eliminates the need for device testing.

• Short silence: If a user pauses mid-sentence, a strict silence threshold may terminate the turn before they finish. Test natural hesitations and long numbers.
• Barge-in: Interrupt while the assistant is speaking, then correct an earlier instruction. The assistant must respect the new direction, not complete a stale action.
• WebRTC/SIP behavior: OpenAI documents server-managed output buffering and automatic truncation of unplayed speech during user interruptions.
• WebSocket behavior: Immediately halt local playback, discard queued audio and send conversation.item.truncate with the duration actually played. Otherwise the model may retain words the user never heard.

SXF analysis: interruption is a state-consistency problem, not only a response-speed metric. An agent can sound responsive yet issue an outdated tool call. Measure both the audible conversation and the authorized final state. Official VAD reference ↗

ORIGINAL SXF METHODOLOGY · VERSION 1.0

SXF Voice Agent Evaluation Lab: seven controlled tests

We designed a reproducible seven-scenario scorecard to turn “the voice sounds good” into inspectable evidence. This is a published protocol and unfilled scorecard; we have not performed these paid model calls, and there are no measured pass rates, latency percentiles or winners in this article.

How to publish credible results later

1. Freeze exact source audio, instructions, model ID, model settings, voice, transport and microphone chain.
2. Run each scenario more than once and vary the seed audio across independent participants; repeated runs on one fixture are not an estimate of general performance.
3. Log model usage, tool calls and event timestamps. Keep incorrect output rather than deleting failures.
4. Score separately: semantic correctness, spoken identity accuracy, task acceptance, unauthorized effects, first-useful-audio time, interruption recovery and connection reliability.
5. Publish sample size, errors, uncertainty and hardware limitations. Never present vendor launch claims as independent SXF results.

This aligns with the reliability framework in SXF’s twelve agent reliability metrics and the 25-test agent evaluation checklist.

VERIFIED PRICING · NOVEMBER 2026

GPT-Realtime-2.1 API pricing: audio vs text vs image tokens

The published Standard API prices are per million tokens by modality. A five-minute call does not carry a universal fixed price. Audio duration, silence, outputs, history/caching, thinking and tool usage can affect the tokens billed.

For the exact price book, rate tiers and billing caveats see OpenAI API pricing. Do not confuse these rates with gpt-live-1, which has its own session-based billing model. Also do not confuse cached input with ordinary input; the calculator below defaults to non-cached input.

ORIGINAL SXF TOOL · CLIENT-SIDE ONLY

Realtime voice API cost calculator: model vs Mini

Use actual token counts from your usage logs—or change these clearly hypothetical inputs. No input is sent to a server. The calculator compares published list prices and excludes taxes, carrier charges, additional tools, discounts, retries and other products.

Interpretation: Mini’s lower list price is not the same as lower cost per accepted outcome. If a cheaper model fails more tasks, retries or escalations can outweigh raw token savings. Measure the denominator: successfully completed, authorized business work. See SXF’s enterprise agent TCO research.

MODEL SELECTION · SEARCH COMPARISON

GPT-Realtime-2.1 vs GPT-Realtime-2.1 Mini: which should you test first?

Decision rule: test both on the same seven scenario IDs, then compute cost per accepted task = total verified API + tool + operating cost ÷ accepted task count. If no accepted tasks were observed, the metric is undefined—not zero.

Open the official Mini model card and cross-check SXF’s pricing database. If you need telephony, CRM integration or business ROI, continue to the AI receptionist build guide.

DIAGNOSTICS · PRACTICAL FIXES

GPT-Realtime-2.1 Playground not working? Seven checks

DIRECT QUESTIONS

GPT-Realtime-2.1 Playground FAQ

EVIDENCE POLICY · SXF RESEARCH METHOD

Primary sources, analysis boundaries and update policy

Source-verified: model specifications, list prices, session transport, VAD options and interruption semantics link to first-party OpenAI API documentation. Original SXF contribution: the seven-scenario experimental design, unfilled evaluation dataset, interpretation of failure modes and deterministic token-cost calculator.

Not measured here: independent model accuracy, real-world latency, speaker intelligibility, tool completion rates or average cost per minute. We have not run the Playground on behalf of site visitors. Worked token figures are stated assumptions, not vendor telemetry.

1. OpenAI · GPT-Realtime-2.1 official model card — capabilities, model ID, limits and Playground entry.
2. OpenAI · GPT-Realtime-2.1 Mini model card — comparison model and supported modalities.
3. OpenAI · API pricing — audio, text and image per-token pricing.
4. OpenAI · Getting started with Realtime API — client secrets, SDK and transport.
5. OpenAI · Voice activity detection — VAD modes and speech events.
6. OpenAI · Realtime conversations — interruption and truncation behavior.
7. OpenAI · Voice agent evaluation — task outcome, instrumentation and repeatability principles.

Primary-source facts reviewed November 8, 2026. Pricing and interface controls can change; verify current official values before a production decision. For editorial policies, see About SXF / AI.