Release notes

Release notes

Changes to the video APIs and their models.

2026-10-17 Preview

  • Realtime sessions API: POST/GET/DELETE /v1/realtime/sessions, per-model command schema, and an authenticated WebRTC signaling proxy.
  • Queueing with position and estimate, one session per GPU pod, retry on a second pod if the first fails to start.
  • Per-second metering from ready; billing stops at the last healthy check if a pod dies.
  • Models: mock-stream (CPU test model) and longlive-1.3b (evaluation only; not yet measured on production GPUs).
  • Reference Python viewer that negotiates WebRTC through the proxy and reports time to first frame, frame rate, stalls, and whether a prompt change is visible.
  • Known gaps: no browser SDK, no short-lived client tokens, no moderation, no per-account quotas, no published pricing, and media reachability from rented GPUs is unverified.