OpenAI's Realtime API v2 beta cuts voice latency ~25%

Alongside GPT-5.6's staged rollout, OpenAI shipped a Realtime API v2 beta focused squarely on voice agents. The update reduces voice-to-text latency by roughly 25% versus v1 and adds streaming function calls and interrupt handling — the ingredients needed for natural, barge-in-capable voice interactions. Developers cited ~30% lower end-to-end latency, 95%+ interruption success and 40% faster tool execution, targeting sub-100ms customer-support agent workflows.

The mechanism that matters is interrupt handling: real conversations involve people talking over each other, and an agent that can be interrupted mid-response and gracefully recover feels dramatically more natural than turn-based voice bots. Combined with streaming function calls, the API can execute tools (look up an order, issue a refund) while the conversation continues, collapsing the awkward pauses that betray a bot.

OpenAI paired this with a Batch API repricing effective July 8 introducing cheaper overnight job tiers — a cost lever for non-latency-sensitive bulk work, and a small counter to the week's cost-pressure narrative that has Microsoft insourcing and enterprises eyeing Chinese models.

The competitive context is sharp: SpaceXAI's Voice Agent Builder launched at $0.05/min with 80+ voices and cloning, which developers see as a viable lower-cost call-center alternative to OpenAI's Realtime API. Voice agents are becoming a price war as much as a capability race. Watch whether OpenAI's latency lead holds against xAI's pricing, and whether Realtime v2 graduates from beta with pricing that survives the cost scrutiny facing all frontier vendors this week.