OpenAI moves the Realtime API to general availability with SIP calling, MCP, and a cheaper speech-to-speech model

What shipped: one model instead of a pipeline

OpenAI has taken the Realtime API out of the public beta it entered in October 2024 and made it generally available, releasing it with a new speech-to-speech model called gpt-realtime and two exclusive voices, Cedar and Marin. The existing eight voices are being updated to inherit the same quality improvements.

The architectural claim is the anchor of the release: rather than chaining a speech-to-text model to a language model to a text-to-speech model, the Realtime API processes and generates audio directly through a single model and API. OpenAI argues this reduces latency and preserves nuance — non-verbal cues like laughs, mid-sentence language switches, and tone shifts between 'snappy and professional' and 'kind and empathetic' — that a transcription-first pipeline tends to flatten. For teams building voice agents, this is the difference between reconstructing a conversation from text tokens and keeping the audio signal intact end to end.

The benchmark deltas that justify calling it production-ready

OpenAI backs the 'production-ready' framing with three benchmark comparisons against its December 2024 model. On Big Bench Audio, which adapts Big Bench Hard reasoning questions into the audio domain, gpt-realtime scores 82.8% against the prior 65.6%. On an audio version of MultiChallenge — a multi-turn instruction-following test — it scores 30.5% versus 20.6%. On ComplexFuncBench, converted from text prompts to speech, function calling reaches 66.5% against 49.7%.

The MultiChallenge number is worth reading honestly: 30.5% is a meaningful jump over 20.6% but still means the model mishandles the majority of these deliberately hard multi-turn cases. That matters because instruction adherence is precisely where voice automation breaks — reading disclaimer scripts word-for-word, repeating back alphanumerics, or refusing to do something a system prompt forbids. OpenAI also notes improved detection of alphanumeric sequences like phone numbers and VINs in Spanish, Chinese, Japanese, and French, which is the unglamorous accuracy work that decides whether a support call actually completes.

The plumbing that turns a model into an agent: SIP, MCP, and async calls

The most consequential additions for automation are not the voice quality but the integration surfaces. SIP support connects sessions directly to the public phone network, PBX systems, and desk phones — meaning a voice agent can answer or place real calls without a separate telephony layer. Remote MCP server support lets a session point at a server URL (the docs example uses Stripe's hosted MCP endpoint) and have tool calls handled automatically, so swapping the server swaps the agent's capabilities.

Asynchronous function calling addresses a specific failure mode of live voice: a long-running tool call used to stall the conversation. gpt-realtime now continues talking while waiting on results, and OpenAI says the behavior is native, requiring no code changes. Combined with reusable prompts — saved bundles of developer messages, tools, variables, and example turns carried across sessions — these are the components that separate a demo from a deployed agent. Zillow's Josh Weisberg framed the payoff in product terms:

The new speech-to-speech model in OpenAI's Realtime API shows stronger reasoning and more natural speech—allowing it to handle complex, multi-step requests like narrowing listings by lifestyle needs or guiding affordability discussions with tools like our BuyAbility score.Montana Labs

Pricing and context control aimed at long sessions

gpt-realtime is priced 20% below gpt-4o-realtime-preview: $32 per million audio input tokens ($0.40 for cached input) and $64 per million audio output tokens. Alongside the cut, OpenAI added fine-grained conversation-context control that lets developers set token limits and truncate multiple turns at once — a direct lever on the runaway cost of a voice session that never resets.

That combination signals what OpenAI thinks the blocker was. Voice minutes are expensive when every turn accumulates audio tokens, and a call center or assistant that runs for many minutes can become uneconomical fast. Cutting per-token price while giving developers the ability to prune context is a more honest fix than price alone, because it attacks the growth in context length that makes long voice calls costly in the first place.

The specific implication: telephony automation now runs inside the model provider

With SIP calling native to the API, OpenAI has absorbed a layer that voice-agent builders previously stitched together from separate telephony and speech vendors. A team can now attach a phone number, connect remote MCP tools, and run a single speech-to-speech model with async tool calls — without a transcription pipeline in the middle. The safety posture reflects that this is meant for real deployments: active classifiers can halt sessions that violate content rules, only preset voices are allowed to block impersonation, developers must disclose AI to users, and EU data residency is supported.

For applied teams, the practical read is that the build-versus-buy calculus for voice automation just shifted toward buying the whole stack from one API — with the tradeoff that call routing, tools, and voice quality are now coupled to OpenAI's roadmap and its 30.5% ceiling on hard multi-turn instruction following. That number is the one to validate against your own scripts before wiring a live phone line to it.