---
scope: org
organization: livekit
release_count: 184
last_release: 2026-10-15
generated: 2026-04-21
updated: 2026-10-17
canonical: https://releases.sh/livekit/overview.md
---

Recently shipped GPT-Live over LiveKit Inference, Gemini Live and Grok Voice model support, and a JS SDK fix for leaked audio contexts.

**GPT-Live runs on LiveKit Inference.** `AgentSession(llm="openai/gpt-live-1")` uses LiveKit credentials instead of a separate OpenAI key, with `inference.DuplexModel` handling the protocol. Gemini Live model strings and xAI Grok Voice 3.0 joined the Inference list, alongside a TypeSafe Jev decision-model plugin. **Breaking:** `inference.RealtimeModel` no longer subclasses the OpenAI plugin's realtime model, so `isinstance` checks need updating.

**Python agents can measure realtime reply latency.** livekit-agents 1.8.7 reports end-to-end latency — last detected user speech to first assistant playback — from assistant message metrics when local VAD is configured. **Breaking:** telemetry fallback attributes moved to adapter request spans, Anthropic custom HTTP clients must use `httpx2`, and the OpenTelemetry Events API is replaced by `LogRecord(event_name=...)`.

**Rust SDKs can mint agent credentials.** livekit-api 0.9.0, livekit-token, and livekit-uniffi add the `agent` grant and `kind` claim that Go, Python, and JS already carry; **breaking**, the four server-inferred grants are now `Option<bool>`, read through server-matching getters. Data streams error with `AbnormalEnd` when a room closes, a room dropped without `close()` no longer leaks its RTC engine, and livekit-capture 0.1.2 parses H.264/H.265 access units and adds GPU wall-clock and test-pattern sources.

**Egress moved to Chrome 154**, closing CVE-2026-85046 and CVE-2026-87491, two V8 remote-code-execution bugs exploited in the wild; egress runs Chrome unsandboxed unless `enable_chrome_sandbox` is set. Egress also gained a native OCI Object Storage upload backend and SDK-source compositing for room composite recordings.

**Smaller surfaces.** Client SDK JS v2.23.0 releases audio contexts and web audio graphs after `room.disconnect()`, which previously accumulated across mute/unmute cycles. Server v1.13.10 filters VP8, VP9, and AV1 simulcast temporal layers with dependency descriptors; v1.13.9 restored egress-start behavior broken in v1.13.8. `lk simulations` merged into `lk agent simulate` as Agent Simulations.