GPT-Realtime-2: GPT-5 Reasoning for Voice Agents

Voice agents just got a reasoning engine

On May 7, OpenAI shipped three new audio models to its API: GPT-Realtime-2, Realtime-Translate, and Realtime-Whisper. The headline model, GPT-Realtime-2, is the first voice-native model built on GPT-5-class reasoning — meaning your voice agent can now think through multi-step problems while holding a conversation. The announcement post on X pulled over 3.4 million impressions in its first few days, and developer interest is tracking accordingly.

This isn't a minor version bump. OpenAI is making the case that voice agents are ready to move from demo-grade toys to production infrastructure. The pricing, the context window, and the tooling all point in the same direction: they want developers building real businesses on top of this.

What GPT-Realtime-2 actually brings

The original Realtime API (launched in late 2024) gave developers streaming voice-to-voice with GPT-4o. It was fast and surprisingly natural, but it had the reasoning depth of a chatbot — fine for simple Q&A, not great for anything requiring actual thought.

GPT-Realtime-2 changes the equation with several key upgrades:

• GPT-5-class reasoning — The model can chain through complex logic while maintaining a natural voice conversation. Think customer support that actually understands your billing dispute, not just pattern-matches keywords.
• 128K token context window — Four times the context of the original Realtime model. Long conversations, detailed reference documents, and multi-turn reasoning all become practical.
• Native tool calling — The model can invoke external functions mid-conversation: look up an order, check inventory, schedule an appointment, then continue talking. This is the feature that turns a voice chatbot into a voice agent.
• Improved interruption handling — Users can cut in naturally, and the model recovers gracefully. This sounds small, but bad interruption handling is the #1 reason voice bots feel robotic.

Per OpenAI's announcement, GPT-Realtime-2 posted a 15.2% improvement on Big Bench Audio, a benchmark that tests audio understanding across diverse tasks. That's a meaningful jump — Big Bench improvements tend to be incremental, so double-digit gains suggest a real architectural upgrade, not just fine-tuning on the eval set.

The two companion models

OpenAI didn't just ship one model — they shipped a trio designed to work together.

Realtime-Translate

A dedicated real-time translation model that sits alongside GPT-Realtime-2. Rather than asking the main model to handle translation as a side task (which eats reasoning capacity), Realtime-Translate handles language conversion as a specialized pipeline. For companies building multilingual voice agents — customer support across Europe, for example — this separation of concerns matters. You get better translation quality without degrading the primary model's reasoning on the actual task.

Realtime-Whisper

An upgraded speech-to-text model built for the Realtime API pipeline. Whisper has been OpenAI's open-source transcription workhorse since 2022, but this variant is optimized for the low-latency, streaming context of real-time voice interactions. Better transcription accuracy means fewer misunderstood inputs, which compounds into significantly better agent performance over a full conversation.

Pricing: $0.017 per minute

OpenAI set the entry point at $0.017 per minute for real-time voice interactions, according to the launch announcement. To put that in context:

Metric	Cost
Per minute	$0.017
Per hour	~$1.02
1,000 calls at 3 min avg	~$51
10,000 calls at 3 min avg	~$510

For a production voice agent handling, say, 10,000 customer calls per month at an average of 3 minutes each, you're looking at roughly $510/month in model costs. Compare that to the cost of a single human agent ($3,000-5,000/month fully loaded), and the unit economics are obvious — if the quality is there.

The "if" is doing work in that sentence. Pricing is only compelling when the model actually resolves issues. Which brings us to the early production data.

Zillow's 95% call success rate

OpenAI highlighted Zillow as an early deployment partner, citing a 95% call success rate with the new Realtime API models. That's a striking number for voice AI in production. For context, industry benchmarks for traditional IVR systems hover around 60-70% resolution rates, and first-generation AI voice agents typically land in the 75-85% range.

A few caveats worth noting: OpenAI hasn't published what "call success" means in Zillow's context — whether that's successful routing, issue resolution, or customer satisfaction. The definition matters enormously. A 95% routing success rate and a 95% full-resolution rate are very different achievements. We also don't know the complexity distribution of those calls.

Still, even with those caveats, 95% is high enough to suggest that GPT-5-class reasoning genuinely changes what voice agents can handle. Simple FAQ bots didn't need better reasoning. But an agent that needs to understand a lease question, cross-reference listing details, and schedule a showing — that's where reasoning depth pays off.

Why this matters for the voice agent market

The voice AI space has been stuck in a frustrating middle ground. The technology is good enough to demo impressively but not reliable enough to deploy without a human fallback on every call. Several things about this launch suggest OpenAI is trying to push past that barrier:

• Reasoning + voice in one model — Previous approaches required chaining a speech-to-text model, a reasoning LLM, and a text-to-speech model. Each handoff added latency and error surface. A unified model that reasons natively over audio removes those failure points.
• Tool calling as a first-class feature — Voice agents that can only talk are novelties. Voice agents that can look up data, take actions, and confirm results are products. Native tool calling in the Realtime API is what makes the "agent" part real.
• 128K context for long conversations — Customer support calls average 6 minutes, but complex cases run 20-30 minutes. A 128K context window means the model won't lose track of what was discussed 15 minutes ago — a problem that plagued earlier voice models.

My read: this launch is less about any single feature and more about crossing a viability threshold. GPT-Realtime-2 is the first voice model where reasoning, context, and tool use are all good enough simultaneously. Previous models had one or two of those three — having all three is what makes production deployment realistic.

The competitive picture

OpenAI isn't operating in a vacuum. ElevenLabs has built a strong position in voice synthesis and recently launched its own conversational AI features. Google's Gemini 2.5 models support native audio. And a wave of startups — Vapi, Bland AI, Retell — have been building voice agent platforms on top of existing LLMs.

What OpenAI is betting on is that reasoning quality is the bottleneck, not voice quality. Most voice AI platforms already sound decent. The problem is that they say dumb things confidently. By bringing GPT-5-class reasoning directly into the voice pipeline, OpenAI is attacking the actual failure mode rather than polishing the audio fidelity.

That said, OpenAI's voice models remain API-only and closed-source. For companies that need on-premise deployment, fine-tuning control, or want to avoid vendor lock-in, the open-source voice AI stack (Whisper + local LLM + Bark/XTTS) remains relevant even if it's behind on raw capability.

What developers should watch for

A few open questions that will determine whether GPT-Realtime-2 lives up to the launch hype:

• Latency in production — Demo latency and production latency under load are different things. OpenAI hasn't published p50/p99 latency numbers for GPT-Realtime-2 yet. For voice agents, anything above 800ms response time starts feeling unnatural.
• Tool call reliability — Tool calling in text-based models is already imperfect (hallucinated function names, wrong parameter types). Whether tool calling is robust enough for voice — where you can't show the user a retry button — is an open question.
• Cost at scale — $0.017/min sounds cheap, but voice agents handle volume. A mid-size call center doing 100,000 calls/month at 5 minutes average would run $8,500/month in model costs alone, before infrastructure, monitoring, and fallback staffing. Competitive, but not trivially cheap.
• Multilingual performance — Realtime-Translate is a smart architectural choice, but we don't have benchmark data on its quality across language pairs yet. European and Asian language support will be critical for enterprise adoption.

The bottom line

GPT-Realtime-2 is the most capable voice agent model available today, at least on paper. The combination of GPT-5-class reasoning, 128K context, native tool calling, and $0.017/min pricing checks every box that was previously blocking production voice AI deployments.

The real test is what happens over the next few months as developers push it into production at scale. Zillow's 95% success rate is promising, but one partner's results don't make a trend. I think the companies most likely to benefit immediately are those with high call volumes, well-structured backend APIs (for tool calling), and use cases where the reasoning requirements go beyond simple FAQ — think insurance claims, technical support, and financial services.

If you're building voice agents, this is worth evaluating now. If you're waiting for voice AI to "be ready" — the gap between current capability and your requirements just got a lot smaller.