OpenAI drops gpt-realtime and makes Realtime API production-ready

Hooking up a voice AI that sounds human has long felt like assembling a tangled mess of speech-to-text, LLM, and text-to-speech services. Half the latency was spent passing data between servers, and the bot still sounded robotic.

OpenAI just officially launched its Realtime API into general availability, alongside a brand-new flagship speech-to-speech model dubbed gpt-realtime. It brings massive benchmark jumps, native phone support, and a 20% price cut.

Smoother voices and async tools

Unlike traditional voice setups that stitch multiple models together, gpt-realtime processes and generates speech directly in a single pass. This keeps latency ultralow and preserves natural inflections.

OpenAI added two new exclusive voices named Cedar and Marin, while updating its existing eight presets. The model can now register laughter, switch languages mid-sentence, and adapt its cadence based on subtle prompts like speaking with empathy or matching a specific accent.

On raw intelligence, gpt-realtime hit 82.8% on the Big Bench Audio benchmark, up from 65.6% on the previous December release. Its ability to accurately repeat alphanumeric sequences like phone numbers and VINs across languages like Chinese, Japanese, and Spanish also saw huge gains.

Crucially, OpenAI solved the awkward pause problem with asynchronous function calling. When an agent triggers a slow backend task like fetching an account balance, the model can keep chatting fluidly without stalling the line.

Dialing in SIP, MCP, and screenshots

The API updates give the model hands and eyes designed specifically for enterprise deployments.

It now natively supports Session Initiation Protocol, or SIP. That means developers can link gpt-realtime straight into PBX hardware, office desk phones, and standard public telephone networks without clunky custom bridges.

It also supports remote Model Context Protocol servers. Developers can simply feed a server URL into the session config, and the API automatically exposes those external tools to the voice agent.

Multimodal image inputs are also live. Rather than demanding a bandwidth-heavy live video stream, the system accepts screenshots and photos on demand, letting users ask spoken questions about what is on their screen.

Pricing dropped 20% compared to the earlier preview. Audio input costs $32 per million tokens, cached inputs run just $0.40, and audio output is priced at $64 per million tokens.

Why it matters

For builders, this moves voice AI from brittle weekend hackathons to reliable enterprise phone queues. If you run a customer service team or build outbound calling agents, you can now spin up an intelligent agent wired directly to a phone switch and internal APIs in hours rather than months.

For the AI ecosystem, OpenAI is aiming straight at traditional telecom and call center software stacks. By packaging SIP support and MCP tool discovery right into the base API, it eliminates the need for expensive third-party middleware and wrappers that merely glued Twilio to an LLM.

We think opting for static image injections rather than raw video feeds was a very sharp engineering trade-off. It keeps compute bills manageable while covering nearly all real-world visual support tasks. The risk, however, sits in instruction adherence. While the MultiChallenge score jumped to 30.5%, that absolute number shows the model can still wander off script, meaning strict compliance-heavy industries will still need tight guardrails before setting it completely loose on phone lines.