Summary
In its July 27, 2026 update, Vercel added streaming transcription support to AI Gateway, letting applications receive incremental speech-to-text output rather than waiting for a full transcript — a building block for real-time voice agents. The same update cut Python function cold starts roughly in half (from about 2.8s to 1.3s for median-sized functions) by shipping build-time bytecode compilation inside function bundles, and enabled drag-to-deploy from the Vercel homepage with no Git, CLI, or local setup.
What changed
Vercel AI Gateway gained streaming transcription; Python function cold starts dropped from ~2.8s to ~1.3s via build-time bytecode compilation bundled with functions; and a no-setup deploy path was added from the Vercel homepage (announced 2026-07-27).
Why it matters
Streaming transcription in the AI Gateway lowers latency for voice-driven agents built on Vercel, matching the realtime-speech direction of the AI SDK. Faster Python cold starts matter because Python is the default language for AI/ML workloads, and cold-start latency has been a persistent tax on serverless AI functions — halving it makes Vercel more viable for latency-sensitive inference and agent backends.
Evidence excerpt
Python cold starts on Vercel became 2x faster (from 2.8s to 1.3s for median-sized functions) thanks to build-time bytecode compilation now included in function bundles; AI Gateway now supports streaming transcription; and you can deploy from the Vercel homepage without Git, the CLI, or any local setup.