ElevenLabs Changelog

Eleven Realtime is a new family of voice agent models, with v1 Mini as the default and v1 Max reasoning before it answers, keeping the same events and WebSocket transport as OpenAI's Realtime API. Realtime Translation, a WebSocket API for enterprise workspaces, turns speech in one language into speech in another while the speaker is still talking.

Text to Speech and Text to Dialogue requests accept an optional audio_effects object applying filter presets, distance, reverb, background noise, and panning, with a catalog endpoint at GET /v1/audio-effects/catalog. Realtime Speech to Text adds turn_prediction as a commit_strategy value, and new agents are created with call queueing enabled unless platform_settings.queueing_config.enabled is set explicitly.

Updating an agent or creating an agent draft now compiles and validates structured procedures automatically, with invalid procedures returning per-procedure errors as 400 procedure_validation_failed, and the compile procedures endpoint is marked legacy. Merge proposals for agent branches are available in alpha, alongside SDK releases adding agent merge proposal clients and realtime Speech to Text transcript editing.

Eleven v4 and Eleven v4 Turbo are now available, with v4 offering higher-quality, more expressive speech and improved voice cloning across more than 90 languages, and v4 Turbo targeting real-time applications at roughly 100 ms median inference latency. Speech to Text batch and realtime transcription now accept natural-language transcript edit instructions of up to 2,000 characters.

Sora 2 and Sora 2 Pro were removed from the Image & Video model picker ahead of OpenAI's September 24, 2026 API shutdown, so Flows and templates using a Sora node must switch to another video model before they run. Seedance 1.5 Pro is deprecated and stops working on November 11, 2026, with Seedance 2.0 models as the replacement.

Agent prompt configuration now includes enable_parallel_tool_calls (default true) so supported models can execute multiple tools in one turn, and image generation supports the gpt-image-2.5-flare and gpt-image-2.5-sunburst models with up to 10 reference images and 1K/2K/4K output. Also added a cursor-paginated phone number endpoint with provider, outbound, agent, branch, label, and phone number filters, Slack alerting for agents, the gpt-6-astra agent LLM, and a cascade_timeout_seconds setting for Speech Engine.

ElevenAgents now queues callers at their concurrency limit, playing hold audio and emitting queue_status events until they are admitted or time out, with new API methods to upload or delete custom hold audio. Music endpoints and SDKs accept music_v2_5, and agent LLM configuration accepts gemini-3.8-flash.

Scribe v2 Medical, a batch speech recognition model specialized for medical and clinical audio, is now generally available at the same rate as Scribe v2. Pass scribe_v2_medical as model_id on Create transcript; keyterm prompting, entity detection, speaker diarization, and no-verbatim mode work the same way as Scribe v2.

Outbound telephony and batch calls now support optional twilio_machine_detection with results via a new answering_machine_detection webhook event. Added workspace-wide ticket listing, dynamic variable conversation filters, and deprecated allowed_values_dynamic_variable in favor of allowed_values.

ElevenAgents now supports DTMF phone keypad input with configurable timeouts and transcript redaction, and conversation transcripts include attached file details. Music endpoints add optional waveform data, and the JS/Python SDKs allow multiple file attachments per message.

The ElevenLabs CLI v1.0.0 now exposes every API operation as a subcommand with JSON, table, YAML and CSV output, and adds local configuration workflows for ElevenAgents. Procedures are now generally available in ElevenAgents, letting one agent handle distinct tasks without loading every instruction in its system prompt, plus new conversation triage ticket APIs and a context_usage event for realtime clients.

The local ElevenLabs MCP server and MCP player are deprecated in favor of the hosted MCP server at https://api.elevenlabs.io/v1/mcp, which signs in via OAuth and requires no local installation or API key. Both local repositories are archived and will no longer receive updates.

ElevenCreative gains asynchronous Flows APIs for video, image and speech generation with webhook support, plus reusable asset APIs for chaining jobs. ElevenAgents adds concurrency wait queues, lightweight conversation summaries, and topic pagination. Background audio defaults changed: volume now 0.15 (was 0.6) and crossfade_loop defaults to true.

Dubbing v2 is now available via a project-based API supporting over 90 languages with editable JSON transcripts. SDK updates add realtime speech-to-text options and conversation guardrail filters; v2.61.0 removals include OpenAI realtime session types and run_subagent_* variants.

Scribe v2 Realtime now detects entities and emits them in committed_transcript_entities events. ElevenAgents adds per-test tool mocks, translated workflow messages, version filters, a knowledge base system tool, VAD config, WebSocket custom LLMs, and DTMF input redaction. SDK packages add onAgentReasoningResponsePart, logging controls, and more.

Eight new procedure APIs for ElevenAgents, knowledge base bulk dependency checks and deletion, and batch call result export as CSV. The Get WebRTC token response now requires conversation_id in addition to token — a breaking schema change for clients that validate the response shape.

Music Finetunes API adds five endpoints for managing finetunes from audio uploads. Agent voice configuration flips enable_phoneme_tags default from false to true, and the backup LLM cascade timeout drops from 8 to 4 seconds.

ElevenAgents adds per-agent sentiment analysis settings, auto-translate transcripts, and nested agent transfers with push/pop/replace workflow operations. Workspaces now support programmatic service-account creation, and member listing via new API endpoints. The optimize_streaming_latency TTS setting is deprecated and documented as a no-op.

Simulate conversation and stream simulate conversation endpoints are deprecated in favor of Create test and Run tests on agent endpoints. Added music detailed stream endpoint for SSE streaming with composition metadata and optional word timestamps. Agents now support numeric evaluation scoring, conversation reasoning, interruption modes per-tool, and platform billing breakdown. Optional disableAuth for Speech Engine WebSocket clients is available in JavaScript and Python SDKs v2.56.0.

Agent configuration branches now support merge and rebase preview endpoints to inspect changes before committing, plus branch-level call volume metrics. Speech-to-text multichannel conversion adds a combined output mode for a single sorted transcript across all channels.