OpenAI Realtime API voice models reset the AI voice agent market

OpenAI Realtime API voice models just shipped a generation jump with GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper, and the price-per-minute on translation is the headline that should worry any voice startup with a moat made of latency. OpenAI announced the three models on 7 May, exiting the Realtime API from beta to general availability on the same day.

Why the OpenAI Realtime API voice models change the maths

The OpenAI Realtime API voice models matter because the audio AI market spent two years building voice agents on top of cascaded ASR-LLM-TTS pipelines. Those pipelines added 700 to 1200ms of latency per turn and demanded a small army of fine-tuned speech models. GPT-Realtime-2 collapses that into a single endpoint that listens, reasons, calls tools, optionally produces an image input response and answers in voice, all without leaving the model. That is the architectural step Google’s Gemini Live and Anthropic’s Claude voice tier have been working toward, and OpenAI got there first to general availability.

Pricing is the second part of the story. £0 (about $0.03)4 per minute for live translation across 70-plus input languages is below Google Cloud Translate’s streaming tier and below ElevenLabs’ multilingual offering. We covered the ElevenLabs AI music push last month, and the same company is now under genuine pricing pressure on its core voice business. Whisper as a paid streaming product at £0 (about $0.01)7 per minute also kills the calculus of self-hosting Whisper-3 for anything below several thousand hours per month.

GPT-Realtime-2 reasoning levels and tool calling

GPT-Realtime-2 is the first speech-to-speech model OpenAI describes as having GPT-5 class reasoning. Developers can pick from minimal, low, medium, high and xhigh reasoning effort, with low set as the default. Low is the sensible balance for call-centre work, restaurant ordering, hands-free admin and other latency-sensitive jobs. High is where the model thinks before it speaks, useful for legal triage and any voice agent that needs to weigh options before committing to an action. xhigh is the deep-reasoning lane that OpenAI is positioning for diagnostic and decision-support voice products.

The context window jumped from 32K to 128K tokens, which is the change voice agent builders have been begging for. A 128K window means a single Realtime session can carry an entire two-hour customer history, multiple tool schemas, an MCP server’s worth of memory and the in-call audio without summarisation tricks. Performance benchmarks support the story – OpenAI says GPT-Realtime-2 scores 15.2% higher on Big Bench Audio and 13.8% higher on Audio MultiChallenge than GPT-Realtime-1.5. Those are real margins, not vendor noise.

Translate and Whisper: the OpenAI Realtime API voice models for everyone else

GPT-Realtime-Translate is the model that should make UK and EU contact-centre operators rethink their localisation budgets. 70-plus input languages with 13 output targets covers the European and Asian markets a London-based business actually serves, and at 3.4p per minute it is now cheaper than the legacy outsourced live-translation services in finance and aviation. The catch is target-language coverage – Welsh, Irish Gaelic and Scottish Gaelic are not on the 13-output list yet, and OpenAI has been opaque about when smaller languages get parity. The Anthropic AWS £79 (about $100)B deal we wrote about in April was partly about exactly this kind of voice-translation workload, and OpenAI just made it sharper.

What UK developers should watch

UK developers should care about three things. First, the Realtime API now supports SIP phone calling and remote MCP servers, which means a voice agent can answer a real PSTN number in minutes rather than the weeks Twilio integrations used to demand. Second, image inputs landed at the same time – your voice agent can accept a photo mid-call and act on it, which closes the gap with Google’s Gemini 3.1 Flash Lite multimodal route. Third, the pricing is denominated in USD, so any UK procurement model needs an FX buffer baked in.

The risk for OpenAI is the same as ever – reliability. We wrote about the Claude double outage in April and the OpenAI Realtime API’s beta period had its own service interruptions in Q1. General availability raises the SLA bar, and any UK-regulated voice product needs to plan for fallback ASR. The model is excellent. The dependency is real.

OpenAI Realtime API voice models and the multimodal angle

The launch ships image inputs alongside the new audio models, which closes a real gap. A voice agent on a customer support line can now accept a photo of a damaged product mid-call and respond with a description, an action recommendation and a tool call – all from the same Realtime session. The cascaded alternative would have required a separate vision model, a separate state-passing layer and at least one network hop between systems. OpenAI’s bundled approach cuts that to a single endpoint and a single billing meter, which is exactly the simplification the production voice agent market has been demanding since 2024.

The MCP integration is the other quietly important detail. Remote MCP server support means a voice agent can connect to a corporate knowledge base, a CRM, a calendar and a payment processor without rebuilding the function-calling layer for each. Anthropic’s MCP rollout passed 97 million installs last month, which we covered in our MCP adoption analysis, and OpenAI adopting the standard for Realtime API is the moment the protocol becomes the de facto inter-model integration layer. UK voice product teams that already shipped an MCP server for chat now get voice for free.