변경 로그

October 26, 2026

Fast LLM service tier

ElevenAgents can now run supported LLMs on a fast service tier, which trades a per-token price premium for higher token throughput and lower latency. Set service_tier to fast in the agent’s prompt configuration with Update agent, and check which tiers a model supports in the available_service_tiers field returned by List LLMs. The default remains standard, and the LLM usage calculator accepts service_tier to estimate the cost difference.

Knowledge base crawl exclusions

ElevenAgents knowledge base website crawls can now skip parts of a site. Pass up to 50 exclude_patterns to Create crawl job, or fill in the Exclude patterns field in the whole-website crawl dialog, and URLs that match any pattern are skipped even when they match the crawl pattern. Crawl job responses return the patterns that were used. See Knowledge base for crawl setup.

Clone strength for v4 voices

Text to Speech and Text to Dialogue requests now accept clone_strength, a value from 0 to 1 that controls how closely a Professional Voice Clone fine-tuned for a v4 model follows its training samples. Set it in voice_settings on Create speech or in settings on Create dialogue. The recommended value for most use cases is 0. On the Text to Dialogue endpoints, clone_strength now decides whether a v4 PVC runs as a PVC or an IVC, and use_pvc_as_ivc is deprecated.

Expiring service account API keys

Workspaces service account API keys can now expire. Pass expires_at_unix when you create or update a key, or pick an Expire after value in the dashboard, and the key stops working after that time. Expiries must be between 15 minutes and 90 days in the future; omit the field for a key that never expires. See Service accounts.

All changes

ElevenAgents

Fast LLM service tier: Agents and workflow overrides accept service_tier (standard or fast) on supported models. List LLMs returns available_service_tiers for each model, and the LLM usage calculator accepts service_tier.
Knowledge base crawl exclusions: Create crawl job accepts exclude_patterns, and the whole-website crawl dialog adds an Exclude patterns field. Matching URLs are skipped.
Procedures analytics: The Procedures tab on the agent analytics dashboard is available to every workspace. Its Procedure invocations chart shows how often each procedure runs over time, grouped by procedure or by agent, and its tooltip shows whether a procedure is structured or free-form.
Start procedures at conversation start: Conversation initiation client data accepts start_procedure_ids, up to five free-form procedures that start when the conversation starts. The agent must have enable_procedure_ids_from_client enabled; structured procedures and unknown IDs fail the conversation start.
Structured procedure steps: Structured procedures can use the end_procedure system tool as a step, and the End step is relabeled End conversation. Sub-procedure and end steps now compile to system tool nodes, so agents start and end the intended procedure more reliably.
Channel integrations out of alpha: The Slack, Telegram, Custom Channel and Intercom integrations are no longer in alpha. Freshdesk remains in alpha.
Slack thread replies: The Slack Channel Message trigger adds a Respond to setting (response_mode), so an agent can keep replying in a thread after a single mention. Mentioning the agent with only stop, mute or quiet silences it in that thread until it is mentioned again, and events that Slack delivers twice get one reply.
Test run names: Run tests accepts an optional name, which test invocations and their summaries return. PATCH /v1/convai/test-invocations/{test_invocation_id} renames a run, and setting name to null restores the generated title.
Branch descriptions: Branch descriptions can be edited from the branches table with Update branch (description, up to 4,096 characters) and appear in the version tooltip and version graph.
Agent duplication keeps traffic splits: Duplicating an agent copies its live traffic split across the copied branches. The resulting deployment records duplicate as its deployment source.
Delete environment variables: DELETE /v1/convai/environment-variables/{env_var_id} deletes an environment variable.
Widget fullscreen on mobile: Widget configuration adds mobile_fullscreen (default true), with a Fullscreen on mobile switch in the dashboard. Set it to false to keep the compact layout on phones. See Widget.
Issues: Triage is renamed Issues in the agents dashboard, and triage tickets are now called issues. Every QA comment and flag_issue_for_review call opens its own issue, and an issue can be reopened while the conversation has another open issue. The API endpoints keep their triage-tickets paths. See Triage.
Prompt caching guidance: The system prompt editor flags system variables that change on every turn and prevent prompt caching, and Spotlight adds an Improve prompt caching recommendation for affected agents.
Realtime voice mode: Switching an agent to Realtime voice mode in the dashboard now selects Eleven Realtime v1 Max. Realtime agents always run tool calls in parallel, so the Parallel tool calling toggle is removed from Realtime model settings. See Realtime models.
Analytics: Workflow analytics can be filtered by channel, and the Guardrails tab is shown to all users.
Analysis LLMs: claude-sonnet-5-5 can be selected as the analysis LLM for evaluation criteria and data collection, both for the agent and per item. Agents on Claude Sonnet 5.5 can also set reasoning effort to none.
Conversation cost: Conversation charging adds voice_surcharge_charge, the credits charged as the voice actor reward for the Voice Library voices a conversation used. It is not included in cost.
Call duration and billing: Call charges and stored call duration stop when the call ends, so teardown time after the call is no longer billed or counted. Calls no longer continue for up to 50 seconds after ending when a knowledge base lookup is in progress.
Telephony: SIP trunk calls forward the User-To-User (UUI) header to the conversation initiation webhook. API keys with an IP allowlist now authenticate on the Genesys, Avaya, AudioCodes and Amazon Connect connector WebSockets. Batch call recipients whose call cannot be placed are marked failed instead of staying dispatched.
Observability: OpenTelemetry post-call webhook agent turn spans add elevenlabs.rag.chunks and elevenlabs.rag.used_chunk_ids, matching the JSON post-call webhook. See OpenTelemetry traces.
Guardrails: In voice conversations, the prompt injection guardrail now evaluates the user turn being answered instead of the agent’s previous message. Conversations on asynchronous channels such as Zendesk, Intercom and Custom Channel that a guardrail blocks now report the guardrail as the termination reason.
Zero Retention Mode: Synthetic voice detection is skipped for conversations in Zero Retention Mode.
WhatsApp: WhatsApp conversations resumed by a reply to a batch call template keep their batch_call metadata and batch ID.
Phone number secrets: Workspace secrets used as a Twilio auth token or an Exotel API token on a phone number show as in use and cannot be deleted while the number depends on them.
Fixes: Text conversations end as soon as the end_call tool runs instead of waiting for the next user message. Voice agents no longer go silent when a turn is cancelled before any audio is sent. The soft timeout no longer fires right after the first message or after tool calls on Eleven v4 voices. Realtime agents with sub-agent, agent tool or prompt say nodes that inherit the LLM can be saved and published. Agent tools that use OAuth2 connections no longer fail with expired tokens when the provider rotates refresh tokens. The nemotron-35-super-vl-nvfp4 context limit is raised from 64,000 to 131,072 tokens.

Text to Speech and Text to Dialogue

Clone strength: clone_strength (0 to 1) in Text to Speech voice_settings and Text to Dialogue settings controls how closely a Professional Voice Clone fine-tuned for a v4 model follows its samples. The recommended value for most use cases is 0. Edit voice settings stores clone_strength and Get voice settings returns it.
PVC and IVC on v4 dialogue: On the Text to Dialogue endpoints, clone_strength now decides whether a v4 PVC runs as a PVC or an IVC, and use_pvc_as_ivc is deprecated.

Voice Changer

New endpoint paths: Voice Changer is documented at POST /v1/voice-changer/{voice_id} and POST /v1/voice-changer/{voice_id}/stream. The /v1/speech-to-speech/{voice_id} paths keep working but are deprecated and removed from the API reference. The Python and JavaScript SDKs call the new paths from v2.75.0, and the SDK method names are unchanged for now. In the next major SDK version, client.speechToSpeech will be deprecated in favor of client.voiceChanger to match the endpoint. See Voice changer.

Workspaces

Expiring service account API keys: Service account API keys accept expires_at_unix on create and update, between 15 minutes and 90 days in the future. Later expiries are rejected with 400 invalid_api_key_expiry. On update, pass "clear" to remove an expiry.
Group concurrency limits enforced: The workspace group concurrency limits (group_concurrency_limit, group_dubbing_concurrency_limit and group_music_concurrency_limit) are enforced for Text to Speech, Voice Changer, Music and Dubbing requests.
Voice cloning permissions: Members without the Create instant voice clone or Create professional voice clone permission can no longer create clones, even with general voice access.
SCIM group seat types: Workspaces that use SCIM without SCIM routing can assign a seat type to SCIM-synced groups from the Groups page.
Fixes: Delete invite returns 404 invite_not_found for an invite that does not exist, instead of a 500 error.

Speech to Text

Shorter realtime silence thresholds: Realtime Speech to Text honors vad_silence_threshold_secs values below 0.5 seconds, which were previously treated as 0.5. The minimum accepted value is lowered from 0.3 to about 0.19 seconds.
Webhook retries: Transcription and error webhooks are retried on failure when the webhook’s retry setting is on. See Speech to Text webhooks.

Dubbing

Legacy endpoints deprecated: The /v1/dubbing endpoints outside /v1/dubbing/project, including Dub a video or an audio file and List dubs, are marked deprecated in the API reference.
Transcript downloads: The dubbing project menu can download target language transcripts as JSON, and as WebVTT subtitles for recent dubs.

Studio

Media folders: Workspace library files can be organized into folders in the Studio media browser.
Video clip actions: Video clips offer voice changer and voice isolator actions.
Video to music: Studio’s video to music workflow uses Music v2.5 by default.

Image and video

Retired models: Image & Video and Flows requests for a retired model fail before generation with a 410 model_retired error that names the replacement model. Flows preflight reports the error on the affected node.
Fixes: Generations wait for a just-uploaded reference asset to finish processing instead of failing.

Voices

Fixes: Cantonese Professional Voice Clone samples are no longer rejected when language detection reports Chinese, and adding PVC samples no longer fails with a 500 error when language detection is unavailable.

Billing

Minimum charge: Text to Speech and Speech to Text requests priced below one credit are charged one credit, and refunds return exactly the credits that were charged.
Business plan image and video pricing: Business plans on the current pricing pay the same image and video credit rate as other self-serve plans, about 17% fewer credits per image and per video second than before.

Realtime Translation

OAuth scope: OAuth clients can request the realtime_translation scope.

SDK releases

JavaScript SDK

v2.75.0 - client.speechToSpeech.convert and client.speechToSpeech.stream now call POST /v1/voice-changer/{voice_id} and its /stream path. Added client.environmentVariables.delete, client.conversationalAi.tests.invocations.update for renaming a test invocation, and an optional name on client.conversationalAi.agents.runTests. Also adds serviceTier on the LLM usage calculators, excludePatterns on knowledge base crawl jobs, expiresAtUnix on service account API keys, cloneStrength in voice settings, startProcedureIds in conversation initiation client data and the widget mobileFullscreen setting. The legacy dubbing list, create, get and delete methods are marked deprecated, and the branch rebase request type BranchesRebaseRequest is renamed to BodyRebaseABranchOntoMainV1ConvaiAgentsAgentIdBranchesBranchIdRebasePost.

Python SDK

v2.75.0 - Added the callback_disconnect option to Conversation and AsyncConversation. It receives a DisconnectionDetails with a reason of "user", "agent" or "error", plus the WebSocket close code and reason when available, and replaces the deprecated callback_end_session. The regenerated client moves client.speech_to_speech to the /v1/voice-changer paths and adds client.environment_variables.delete, client.conversational_ai.tests.invocations.update and the same new parameters as JavaScript SDK v2.75.0.

Android SDK

v0.13.0 - agent_tool_request is now handled as a notification event instead of a client tool call, so client tools no longer run twice, the first time without parameters. The SDK also no longer crashes when a client tool answers after the connection has closed.

ElevenLabs CLI

v1.7.1 - say uses eleven_v4 and new agent templates use eleven_v4_turbo by default. Fixes onboard sign-in on Windows and when the token cannot be saved to the keychain.

Packages

@elevenlabs/client@1.29.0 and @elevenlabs/react@1.19.0 - Added the onAudioAlignmentPlayback callback, called with { event_id, alignment } when an aligned audio chunk starts playing, for syncing text with speech. It is not called on WebRTC connections.
@elevenlabs/client@1.30.0 - Added the opt-in stripAudioTags session option, which removes audio tags such as [laughs] from agent messages passed to onMessage and onAgentResponseCorrection, and exports the stripAudioTags helper. Stripped text no longer lines up with onAudioAlignment timings.
@elevenlabs/client@1.31.0, @elevenlabs/types@0.26.0 and @elevenlabs/react@1.20.0 - Added overrides.tts.modelId for a per-conversation TTS model override. Realtime Speech to Text emits RealtimeEvents.AUDIO_STARTED once the first microphone audio frame is sent, and useScribe in microphone mode reports connected only from that point.
@elevenlabs/convai-widget-core@0.21.0 and @elevenlabs/convai-widget-embed@0.21.0 - Voice call transcripts are revealed in step with the agent’s speech, and an interrupted reply keeps only what was said. Resuming a persistent text conversation no longer asks the user to accept the terms again.
@elevenlabs/convai-widget-core@0.21.1, 0.21.3 and 0.21.4 - Text chat shows typing dots while the agent prepares its reply. Rich content carousels show a page indicator, stay clickable after the conversation moves on, and appear in voice calls while the agent is still speaking. On screens narrower than 768px the chat opens fullscreen; add mobile-fullscreen="false" to the embed to keep the compact layout.
@elevenlabs/react-native@1.2.31, 1.2.32 and 1.2.33 - Updated to the latest @elevenlabs/client and @elevenlabs/react releases.

API changes

New endpoints

ElevenAgents

DELETE /v1/convai/environment-variables/{env_var_id} - Delete an environment variable. Returns 204.
PATCH /v1/convai/test-invocations/{test_invocation_id} - Rename a test invocation with required name (string, 1 to 256 characters, or null to restore the generated title).

Voice Changer

Voice changer - POST /v1/voice-changer/{voice_id} and Voice changer stream - POST /v1/voice-changer/{voice_id}/stream - The documented paths for Voice Changer. POST /v1/speech-to-speech/{voice_id} and POST /v1/speech-to-speech/{voice_id}/stream are removed from the API reference and deprecated.

Updated endpoints

ElevenAgents

Create agent, Get agent and Update agent
- Prompt configuration adds service_tier (standard or fast, default standard). Workflow node prompt overrides add nullable service_tier.
- Widget configuration adds mobile_fullscreen (boolean, default true).
List LLMs - Each model adds required available_service_tiers (array of standard or fast).
Calculate LLM usage and Calculate expected LLM usage for an agent - Added optional service_tier.
Get widget - Widget configuration adds mobile_fullscreen (boolean, default true).
Create crawl job - Added optional exclude_patterns (array of strings, up to 50 items of up to 2,048 characters each).
Get crawl job and List crawl jobs - Crawl jobs add exclude_patterns (array of strings).
Knowledge base document responses - auto_sync_info removes the deprecated auto_discover field.
Run tests - Added optional name (string, 1 to 256 characters). Get test invocation and List test invocations return nullable name.
Update branch - Added optional description (string, up to 4,096 characters). Pass an empty string to clear it.
Get conversation details
- Conversation charging adds nullable voice_surcharge_charge (integer).
- Agent definition source enum adds architect.
Agent deployments (GET /v1/convai/agents/{agent_id}/deployments) - Deployment source enum adds duplicate.
Conversation initiation client data, including the conversation_initiation_client_data WebSocket event and Submit batch call - Added optional start_procedure_ids (array of strings).
Tool type enum removes api_integration_mcp.

Text to Speech and Text to Dialogue

Create dialogue, Stream dialogue and their timestamp variants - use_pvc_as_ivc is deprecated.
Text to Speech voice_settings and Text to Dialogue settings - Added nullable clone_strength (number, 0 to 1). Get voice settings, Get voice and List voices return it.

Dubbing

List dubs, Dub a video or an audio file, GET /v1/dubbing/{dubbing_id}, DELETE /v1/dubbing/{dubbing_id}, GET /v1/dubbing/{dubbing_id}/audio/{language_code} and GET /v1/dubbing/{dubbing_id}/transcripts/{language_code}/format/{format_type} - Marked deprecated.

Workspaces

Create service account API key - Added optional expires_at_unix (integer or null).
Update service account API key - Added optional expires_at_unix (integer, or "clear" to remove the expiry; omit to leave it unchanged).
List service accounts and List service account API keys - API keys add nullable expires_at_unix (integer).

October 19, 2026

Eleven Realtime

Eleven Realtime is a new family of models for ElevenAgents voice agents, rolling out gradually across workspaces. Each one is a complete voice agent configuration. You pick one model, and ElevenLabs handles the LLM, Speech to Text and Text to Speech settings behind it. Your agent’s system prompt, voice, tools and knowledge base work the same as before.

There are two to choose from. Eleven Realtime v1 Mini (eleven_realtime_v1_mini) is the default, and it’s the fastest and cheapest. Eleven Realtime v1 Max (eleven_realtime_v1_max) reasons before it answers, so use it when your agent has complex instructions or multi-step tool calls.

If you’ve already built on OpenAI’s Realtime API, you don’t need to rewrite your client. The events and the WebSocket transport are the same, so you change the endpoint and the session configuration. Your backend calls Create realtime client secret to get a short-lived secret, and the browser uses it to connect directly.

Realtime models

Choose a Realtime model and configure an agent’s voice mode.

OpenAI Realtime API compatibility

Connect an existing OpenAI Realtime client to ElevenLabs.

Create realtime client secret

Review the request and response for minting client secrets.

Realtime Translation

Realtime Translation turns speech in one language into speech in another while the speaker is still talking. It’s a WebSocket API, available now for enterprise workspaces.

You open a connection with a target_language and a voice_id, then stream audio in. You get translated speech back in that voice. You don’t have to say which language the speaker is using, because it’s detected automatically, but you can set source_language if you know it. Along the way, the server also sends live transcripts of the original speech and the translated text, for example to show captions.

eagerness controls the tradeoff between speed and quality. high starts translating soonest, and low waits for more of the sentence. The translated voice uses Eleven v4 Turbo by default, and you can switch to eleven_v3 or eleven_flash_v2_5 with tts_model_id. For sensitive conversations, set enable_zrm=true to use Zero Retention Mode.

Realtime Translation WebSocket

Review the connection parameters and message types.

Models

Compare the TTS models available for translated speech.

Zero Retention Mode

Keep transcripts and translations out of logs and history.

Video to Music

Video to Music writes a soundtrack for your video. The Video to music API is now available: you upload a video and get back music that fits it, without writing a music prompt from scratch. Use it to score product videos, ads and social clips.

You can send up to 10 video files (200 MB combined). They’re joined in order and scored as one video. To steer the result, add a text description and up to 10 style tags.

Three options are worth knowing. use_audio_as_reference takes the genre, BPM, instruments and mood from the video’s own audio. add_sfx adds sound effects that match the video. separate_tracks returns the music and the sound effects as separate files in a ZIP, so you can mix them yourself.

Video to music API reference

Review the request parameters and response format.

Music overview

Learn what Eleven Music can generate and how it can be used.

ElevenAgents

Alerts: Alerts notify a workspace when agent error rates exceed a threshold or rise sharply above their recent baseline. Monitors cover conversation failures, tool failures, webhook errors, guardrail triggers, call initiation errors and evaluation success scores, and run with default thresholds without any setup. Agent-level thresholds and notifiers can be set through platform_settings.alerting on Update agent. Notifications can be delivered as webhooks, to Slack or to PagerDuty.
Guardrails generally available: Guardrails are no longer in alpha.
Workspace environments: Environments for environment variables can now be managed through the API with List environments, Create environment and Delete environment. A variable without a value for an environment falls back to its production value. Deleting an environment removes its values from all variables, and production cannot be deleted.
Answering machine detection actions: Twilio answering machine detection settings on Outbound call and Submit batch call accept actions, a list of rules that pair an action (hang_up) with the answered_by verdicts that trigger it, such as machine_start or fax. Verdicts without a rule leave the call running.
New LLMs: Agents can use gemini-4-argon, claude-opus-4-6 and the ElevenLabs-hosted nemotron-35-super-vl-nvfp4. See Models.
Authored-only workflow reads: Get agent and Update agent accept include_procedure_generated_workflow_elements. Set it to false to receive only the workflow nodes and edges you authored, excluding the ones compiled from structured procedures. Defaults to true, so existing callers are unchanged.
Branch rebasing from any branch: Rebase branch and Preview rebase accept an optional source_branch_id to pull changes from a branch other than main.
Delete agents used as subagents: Delete agent accepts force. When true, the agent is deleted even if other agents use it as a subagent, and it is removed from those agents and their branches.
Sort agents by last update: List agents accepts sort_by=updated_at, and each agent summary includes updated_at_unix_secs, the last time the main branch or agent-level settings changed.
Google Calendar event updates: The Google Calendar integration adds google_calendar_update_event and google_calendar_delete_event tools. Both are privileged tools that can change or cancel any event on the connected calendar.

Music and Sound Effects

Zero Retention Mode: Music generation, composition plan, stem separation and video to music endpoints, plus Sound Effects, accept enable_logging=false to use Zero Retention Mode. Zero Retention Mode is available to enterprise customers. Music requests that set store_for_inpainting=true are rejected in this mode.
WAV output for sound effects: Sound Effects accepts wav_8000, wav_16000, wav_22050, wav_24000, wav_32000, wav_44100 and wav_48000 output formats.

Speech Engine

40-minute conversations: Speech Engine supports conversations of up to 40 minutes. The default limit remains 10 minutes. Raise conversation.max_duration_seconds to allow longer conversations.
Sort by last update: List Speech Engines accepts sort_by=updated_at, and each summary includes updated_at_unix_secs.

Dubbing

Subtitles for dubbed languages: Dubbing project language outputs (GET /v1/dubbing/project/{project_id}/language/{language_id}) include subtitles, a signed URL to the translated lines as WebVTT subtitles timed to the source speech. The URL expires one hour after it is issued.

Voices

Resubmittable voice reviews: Voice sharing review_status adds declined_resubmittable for declined voices that can be resubmitted for review.

Workspaces

Disable service account API keys: Disable service account API key disables a service account API key without changing its other settings. It remains available after a workspace downgrades from a plan that includes service accounts.
Group concurrency limits: List workspace groups and List service accounts return group_concurrency_limit, group_dubbing_concurrency_limit and group_music_concurrency_limit for each group, and group permissions add all_resources_access.