ElevenLabs

Recently shipped resumable text conversations in the embedded widget and persistent-session support across the client SDKs.

Text conversations now survive dropped sockets and reloads. The widget's persistent-session attribute stores a resume token per agent, reopens the chat on load, and replays server history when the sheet reopens or the next message sends. @elevenlabs/client adds an experimental persistentSession option plus an onConversationHistory callback for replayed transcript rows. "End chat" forgets the conversation.

Eleven v4 and v4 Turbo are available. eleven_v4 adds more expressive speech and improved voice cloning across 90+ languages; eleven_v4_turbo serves realtime agents at roughly 100 ms median latency.

Speech to text edits transcripts in natural language. Batch and realtime transcription accept an edit instruction up to 2,000 characters; batch returns edited_transcript, realtime emits an edited_transcript event. Not combinable with entity detection.

Agent operations gained ticket triage, workflow outcomes, and Slack alerts.

• Ticket merging folds up to 50 open tickets from one agent into a target; merged tickets keep a merged_into_ticket_id.
• Ticket lists accept a case-insensitive search and return a search_match excerpt.
• Workflow End nodes record success or failure; failures surface as workflow_end_failure.
• New agents default to call queueing.

Also shipped: structured procedures compile and validate on save — invalid ones return 400 procedure_validation_failed, and the standalone compile endpoint is legacy — plus merge proposals in alpha, GPT Image 2.5, Music v2.5, and Scribe v2 Medical at GA. Sora 2 and Sora 2 Pro retired September 24; Seedance 1.5 Pro deprecates November 11.

Breaking in SDK v3 alphas: conversationalAi became agents and Python's on_prem_config became orchestrator_config; install with npm i @elevenlabs/elevenlabs-js@alpha or pip install --pre elevenlabs.