변경 로그

• Analysis items API: Evaluation criteria and data collection items can be managed as reusable analysis items under /v1/convai/analysis-items, with endpoints to list, create, get, update, copy and delete items, list versions, list dependent agents and update attachments in bulk.
• Noisy environment detection: turn.noisy_environment_detection (default false) and turn.noisy_environment_message are available in the API, along with noisy_environment_message_translation on language presets and noisy_environment_message in conversation and workflow overrides. Available for voice agents in Enterprise workspaces. See Noisy environment detection.
• Eleven Realtime available to all workspaces: Realtime voice mode, the Realtime WebSocket and the Realtime Conversations page are available to every workspace, completing the gradual rollout announced on October 19. Realtime conversation groups can now be managed through the API, under /v1/convai/realtime/conversation-groups.
• Realtime API behavior: Each manual input_audio_buffer.commit transcribes only the speech since the previous commit, and input_audio_buffer.clear drops buffered audio. response.metadata from response.create is echoed on response.created and response.done, with at most 16 string pairs. input_audio_buffer.speech_started and speech_stopped are sent once per committed user turn instead of on every pause. Idle audio sessions close after 3 minutes instead of 60 seconds, after a session_idle_timeout error.
• GPT-6.1 Terra: gpt-6.1-terra can be selected as the agent LLM and as the analysis LLM, on the standard and fast service tiers, with reasoning effort from none to max (default medium).
• Default LLM: New agents created without an explicit LLM use gemini-3.5-flash instead of gemini-2.5-flash, and the default backup LLM cascade leads with Gemini 3.5 Flash models. Existing agents are unchanged. Gemini 3.5 Flash-Lite is also available in the India and Singapore isolated environments.
• Model deprecations: gpt-4.1-nano and gpt-4o-2024-05-13, which OpenAI shut down on October 23, are deprecated immediately, and agents using them run on gpt-5.6-luna and gpt-5.6-sol respectively. gpt-5, gpt-5-mini, gpt-5-nano and their 2025-08-07 snapshots are deprecated ahead of OpenAI’s December 11 shutdown, with traffic moving to gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna in steps from 10% on November 11 to 100% on December 6. List LLMs returns deprecation_info for these models. Claude Sonnet 4.5 is deprecated on November 30, with Claude Sonnet 4.6 as the replacement, and is removed from model selectors. The fallback ramp for gemini-2.5-flash and gemini-2.5-flash-lite now starts on November 8 instead of November 1; the November 30 cutoff is unchanged.
• Conversation message counts: List conversations and Get conversation details return user_message_count and agent_message_count. message_count on the conversation list now counts only turns with text, so tool-only turns are no longer counted. Conversations with no messages get a call_successful value of unknown, and the dashboard shows them as “No conversation”.
• Knowledge base: Update document accepts enable_auto_sync, auto_remove and minimum_frequency_days, and the dashboard can edit these auto-sync settings after a document is created, optionally applying folder settings to existing documents. Sitemap crawls accept exclude_patterns, with an Exclude patterns field in the sitemap step of the URL dialog. Website crawls deduplicate pages by canonical URL, and Google Drive auto-sync pauses when the Drive token is rejected and resumes after you reconnect. See Knowledge base.
• Agent testing: Tool call tests accept guardrail_trigger to assert that a guardrail should or should not trigger on the generated turn, with Should trigger and Should not trigger options in the dashboard. Tests whose turn is blocked by a guardrail now fail with “Agent response blocked by guardrail” instead of returning an inconclusive result. List test invocations accepts created_by_user_id, and the test history adds a creator filter. Test runs are returned in creation order. See Agent testing.
• Branches and merges: Branches can have a default environment, set when creating or publishing to a branch and changed from the branch menu. Conversations on the branch use that environment unless the caller requests another one, and such branches cannot receive live traffic. Merge and rebase conflicts can be resolved with Architect, and procedure conflicts can be edited and combined like prompt conflicts. A single procedure’s unpublished changes can be discarded without discarding the whole draft. Branches whose only version was created with changes can be merged, and branch responses report a current_live_percentage of 100 for the main branch when the agent has no live deployment.
• Workflows: Update state nodes and the update_state tool can override LLM settings (model, temperature, token limit and reasoning) for the rest of the conversation from a new LLM tab. Double-click a workflow edge to add reroute points that shape it. Workflow tool nodes accept only configurable system tools plus start_procedure and end_procedure.
• Issues: Issue descriptions can be edited from the issue details sheet, and Update triage ticket accepts qa_comment. See Triage.
• WhatsApp: WhatsApp conversations record call_result (answered, voicemail, busy, no_answer or failed), and conversations opened with a template record opening_message with the template name, delivery timestamps, error code and pricing category. Call permission requests record responded_at_unix, permanent and granted_by_user_call. WhatsApp batch calls that are declined, missed or cannot be placed mark the recipient failed instead of leaving it initiated. See WhatsApp.
• Widget: Widget configuration adds transcript_speech_sync_enabled (default true), with a Reveal transcript as the agent speaks switch; set it to false to show each message in full at once. show_rich_content carousel cards accept rating (0 to 5) and review_count, shown as stars. Resumed persistent text conversations replay rich content in their history. See Widget.
• Channel integrations: Asynchronous channels such as Slack, Telegram, Intercom, Gmail, Salesforce and Custom Channel record message timestamps from the source platform and store the conversation’s main_language from the language override or the agent’s language instead of always en. Salesforce passes the customer’s email to the agent as integration__salesforce_customer_email, which is not a verified identity. Gmail triggers run on the agent’s current version, back off on Gmail rate limits and set aside threads that keep failing.
• Telephony: Twilio conference transfers can combine post-dial digits with call screening (require_acceptance). Tools that wait for speech to finish, such as transfer_to_number and end_call, are skipped if the caller hangs up during their pre-tool speech. Calling skip_turn after a turn timeout starts the skip-turn wait, so the silence timeout no longer ends the call during it.
• Environment variables: Deleting an environment variable that the tip of a live branch references fails with 409 and lists the affected agents and branches. The environment variables table adds Remove value and Delete row actions.
• Dashboard: The conversation view shows human agent replies captured after an escalation in a separate post-escalation section. The system prompt editor’s wand opens an Ask Architect menu to review and improve the prompt or write a new one. Filtered conversation exports include evaluation criteria and data collection results. Turns where the agent speaks after a silence timeout no longer report end-to-end latency. The agent voice settings show the Clone strength slider for every voice fine-tuned for a v4 model.
• Fixes: Files that the LLM provider cannot process are dropped instead of failing later turns. Deleting an agent removes its trigger connections, so the integration connection can be deleted afterwards. The prompt editor saves & and > as typed instead of encoding them. Say nodes no longer duplicate messages, merge proposals that add a node group no longer fail, and node groups show their configured titles. The structured procedure editor opens procedures authored through the API, and dynamic variables in If step expressions are included in referenced_dynamic_variables. Agents whose data collection item uses a dynamic variable can be published again. Knowledge base rediscovery adds pages to older crawl jobs. The concurrency chart no longer overstates peaks when grouped by agent or branch. Garbled characters in LLM tool call arguments are repaired, and WhatsApp webhooks with resent interactive replies or contacts without a name are accepted instead of failing with 422.

• Realtime entity redaction: The realtime WebSocket accepts entity_redaction (all, a category such as pii, phi or pci, or a list of entity types) and entity_redaction_mode (redacted, entity_type or enumerated_entity_type, the default). Each committed transcript is followed by a redacted_transcript event, and the settings are echoed in session_started. entity_redaction cannot be combined with transcript_edit. See Entity detection.
• Realtime authentication: Bearer tokens passed to the realtime WebSocket in Sec-WebSocket-Protocol require the Speech to Text permission, as the other authentication methods do.
• Usage on empty responses: Empty and silent responses from Create transcript include audio_duration_secs.
• Fixes: Webhook transcription requests, with a file upload or cloud_storage_url, now apply diarization_threshold and seed.

• SFX v3 for everyone: SFX v3 (eleven_text_to_sound_v3) is available to every workspace and is the default model in Sound Effects, Studio and Flows. The sound effects that Video to music adds with add_sfx are generated with SFX v3. See SFX v3.

• Professional Voice Clones on Eleven v4: Professional Voice Clone fine-tuning for Eleven v4 and Eleven v4 Turbo is available to every workspace. See Eleven v4.
• Saved clone strength: Eleven v4 requests for a fine-tuned Professional Voice Clone use the voice’s saved clone_strength when the request sets neither clone_strength nor use_pvc_as_ivc.
• Pronunciation dictionary limit: Applying pronunciation dictionaries can expand the input text to at most 50 times its length plus 10,000 characters. Requests over the limit return 400 pronunciation_dictionary_expansion_too_large.
• Fixes: History items for v4 generations keep clone_strength and speed in their saved settings.

• Eleven v4 recommended: Project settings and the model picker recommend Eleven v4 instead of Multilingual v2.
• Larger media uploads: Pro and higher plans can upload audio and video assets of up to 3 GB, up from 1 GB.
• Word uploads: Audiobook projects can be created from Word .docx files as well as EPUB and PDF.
• Video clip background removal: Video clips have a Remove background action, and right-clicking a clip opens its AI Tools.
• Pronunciations tab: Audiobook projects have a pronunciations sidebar tab, opened from the selection menu or with Mod+Shift+P.
• Fixes: The import speech endpoints for external audio and video clips return 429 system_busy and refund the transcription charge when transcription capacity is unavailable, instead of returning 500. Duplicating a project no longer fails when an image clip points at a source image that is not copied.

• Close codes: The WebSocket always closes with a code: 1000 for a normal end, 1008 for a rejected request such as an unknown voice_id or an authentication failure, and 1011 for a server error.
• Invalid frames: Frames that are not a JSON object receive an input_error message and the session continues.
• Fixes: Translated audio is encoded in the requested output_format, so μ-law and A-law output plays correctly. Repeated utterances are no longer dropped, Chinese, Japanese and Korean clauses are no longer spoken twice or skipped, and Korean word order is improved.

• SVG assets: SVG uploads to Assets stay SVG and render as vectors, including in the image editor. Scripts, event handlers and remote references are removed, models receive a PNG, and asset responses omit preview_content_url for SVGs.
• Flows: Node groups and the structure sidebar are available in every workspace. LLM nodes on video-capable models such as gemini-3.5-flash-llm accept video prompt references.
• Clearer reference errors: Generations whose reference file was refused on upload, or is not decodable audio, fail as a validation error with a clear message instead of an unexpected error.

• Longer conversations: Speech Engine conversations can last up to 2 hours, up from 40 minutes. See Speech Engine.
• Request headers: Speech Engine request header values can reference workspace secrets and are validated in the dashboard. Existing headers must be cleared before the WebSocket URL can change.

• Larger voice selections: Manual voice selection is no longer limited to 10 voices, and link-shared voice collections offer Select all. Shared collection links open in the EU, India and Singapore data residency environments.
• Fixes: The voice verification captcha accepts Serbian readings in Latin or mixed script.

• Higher upload limit: Upload music accepts up to 100 songs per user per 24 hours, up from 20.

• Architect permission: Architect access is controlled by its own workspace permission. PermissionType adds agents_architect and group permissions add architect, so admins can turn off Architect without removing Agents access, and API keys have a separate Architect scope.
• History download permission: Download history items requires the speech_history_read API key permission.
• Fixes: New users whose only invite is to a child workspace that inherits or borrows SSO can sign up with that SSO instead of receiving 400 provider_not_allowed.

• Architect charging: Architect usage is now billed from workspace credits as its own line item, separate from agents’ minutes and LLM cost. Billing is rolling out gradually; the composer in the dashboard shows when it applies to your workspace. See Pricing and usage.
• Realtime pricing: Workspace pricing tables list Realtime API audio (eleven_realtime_v1_audio) and the Eleven Realtime v1 Mini and Max token rates.
• Cached LLM tokens: Cache writes are billed at least at the input rate and cache reads at the vendor’s cached rate. LLM calls for post-call analysis, guardrails and agent tests bill cached prompt tokens at those rates instead of the full input rate. Gemini 2.5 Flash-Lite cache reads are billed at 0.01 instead of 0.
• Preview conversation discount: The discount on conversations started from the agent tester applies only to workspaces on 2022 pricing. On 2026 pricing, charging.dev_discount is false.

• Fixes: OAuth discovery names the host the client connected to, so MCP clients such as Claude Code, Cursor and VS Code can connect to https://api.elevenlabs.io/v1/mcp.

JavaScript SDK

• v2.76.0 - The realtime Speech to Text wrapper accepts entityRedaction. Added client.conversationalAi.analysisItems for managing analysis items, and methods on client.conversationalAi.realtime to list, create, get, update and delete realtime conversation groups and to set or clear a conversation’s group.

Python SDK

• v2.76.0 - The realtime Speech to Text wrapper accepts entity_redaction. Added client.conversational_ai.analysis_items for managing analysis items, and list_conversation_groups, create_conversation_group, get_conversation_group, update_conversation_group, delete_conversation_group, set_conversation_group and clear_conversation_group on client.conversational_ai.realtime.

Android SDK

• v0.13.1 - textOnly is now sent to the server, so text-only sessions run as text conversations instead of voice conversations. The top-level textOnly setting takes precedence over overrides.conversation.textOnly, as in the JavaScript SDK, and setting only the override now selects the WebSocket transport.

Unity SDK

• v0.1.0 - First release of the Unity SDK for voice agents, installed through the Unity Package Manager from a git URL. It requires Unity 6.3 LTS or later and runs in the Editor, desktop players and WebGL, with WebRTC available on WebGL. Conversation.StartSessionAsync starts a session with an agent ID, signed URL or conversation token, and the SDK supports client tools, conversation events, messaging, feedback, device selection and spatial audio through an AudioSource. GettingStarted and QuickStart samples are included.

ElevenLabs CLI

• v1.9.0 - agents workflow diagram draws an agent workflow in the terminal, or prints it as Mermaid with --format mermaid or --format mermaid-md. agents graph shows how agents, tools, tests, knowledge base documents and handoffs connect across the agents in agents.json, and --uses lists the agents that reference a given tool or document. Both commands read local files only. Regenerated against the latest API.
• v1.8.0 - Request User-Agent headers include how the CLI was installed (install/<source>, such as npx, Homebrew or the installer). Pass --telemetry off, or set ELEVENLABS_TELEMETRY=off or DO_NOT_TRACK=1, to drop these tags. Regenerated against the latest API.

Packages

• @elevenlabs/client@1.34.0, @elevenlabs/types@0.28.0 and @elevenlabs/react@1.21.0 - Scribe.connect() and useScribe() accept entityRedaction and entityRedactionMode, and dispatch the redacted_transcript message after each committed transcript. useScribe() adds an onRedactedTranscript callback and sets redactedText on the matching entry in committedTranscripts. stripAudioTags removes audio tags written in any language, not only tags in English letters.
• @elevenlabs/convai-widget-core@0.24.0 - With persistent-session enabled, a voice call can be resumed after a page navigation. The next page shows the conversation so far with “Resume talking?”, and resuming starts a new conversation that continues from those turns, skipping the greeting if the agent allows first message overrides.
• @elevenlabs/client@1.32.0 - Self-hosted worklets (workletPaths) are checked against the installed client version. A worklet copied from another release fails at session start with an error that names the worklet and path. Re-copy the files from @elevenlabs/client/worklets/* when upgrading.
• @elevenlabs/client@1.32.1 - Fixes the start of the agent’s first message playing too fast and too high when the browser cannot use the agent’s sample rate. Voice sessions wait for the output resampler and fail with an error if it cannot be created.
• @elevenlabs/client@1.33.0, @elevenlabs/types@0.27.0 and @elevenlabs/convai-widget-core@0.23.0 - Added a cloneStrength TTS override to the client and an override-clone-strength attribute to the widget. History rows of resumed persistent text conversations can carry rich_content, and the widget replays it.
• @elevenlabs/react@1.20.3 - useScribe with autoConnect keeps a single open session under React StrictMode.
• @elevenlabs/convai-widget-core@0.22.0 - Added the transcript-speech-sync attribute; set it to false to show the agent’s full message at once in voice calls. Carousel cards accept an optional rating from 0 to 5 and review_count.
• @elevenlabs/convai-widget-core@0.21.6 and 0.22.2 - Voice calls show typing dots when the agent takes more than a second to start replying, or “Thinking…” under the avatar when the transcript is off. Partially revealed markdown links show their text instead of [blocked], carousel arrows no longer flicker, and carousel images use a 4:5 ratio.
• @elevenlabs/react@1.20.1, 1.20.2, @elevenlabs/react-native@1.2.34, 1.2.35, 1.2.36, 1.2.37, @elevenlabs/convai-widget-core@0.22.1, 0.24.1 and @elevenlabs/convai-widget-embed@0.21.6 to 0.24.1 - Updated to the latest @elevenlabs/client and widget releases.

New endpoints

ElevenAgents

• Analysis items
  • GET /v1/convai/analysis-items - List analysis items, with optional search, kind (evaluation or data_collection), analysis_item_ids, created_by_user_id, sort_direction, page_size and cursor.
  • POST /v1/convai/analysis-items - Create an analysis item.
  • GET /v1/convai/analysis-items/{analysis_item_id} - Get an analysis item.
  • PATCH /v1/convai/analysis-items/{analysis_item_id} - Update an analysis item.
  • DELETE /v1/convai/analysis-items/{analysis_item_id} - Delete an analysis item. Pass force=true to delete it even when agents use it, removing it from the dependent agents and branches.
  • POST /v1/convai/analysis-items/{analysis_item_id}/copy - Copy an analysis item.
  • GET /v1/convai/analysis-items/{analysis_item_id}/versions - List an analysis item’s versions.
  • GET /v1/convai/analysis-items/{analysis_item_id}/dependent-agents - List the agents that use an analysis item.
  • PATCH /v1/convai/analysis-items/{analysis_item_id}/attachments - Attach or detach an analysis item across agents and branches.
• Realtime conversation groups
  • GET /v1/convai/realtime/conversation-groups and POST /v1/convai/realtime/conversation-groups - List and create conversation groups. Create requires title (string, unique per workspace, up to 64 characters) and accepts optional description (string, up to 500 characters).
  • GET, PATCH and DELETE /v1/convai/realtime/conversation-groups/{conversation_group_id} - Get, update and delete a conversation group.
  • PUT /v1/convai/realtime/conversations/{conversation_id}/conversation-group - Place a conversation in a group, with required conversation_group_id. A conversation belongs to at most one group.
  • DELETE /v1/convai/realtime/conversations/{conversation_id}/conversation-group - Remove a conversation from its group.

Updated endpoints

ElevenAgents

• Create agent and the agent get and update endpoints
  • The prompt llm default changed from gemini-2.5-flash to gemini-3.5-flash.
  • TurnConfig adds noisy_environment_detection (boolean, default false) and noisy_environment_message (string). Language presets add nullable noisy_environment_message_translation, turn overrides and workflow overrides add nullable noisy_environment_message, and the override config adds noisy_environment_message (boolean, default false).
  • Widget configuration adds transcript_speech_sync_enabled (boolean, default true).
• LLM enum adds gpt-6.1-terra, including in conversation_config_override on the /v1/convai/conversation WebSocket.
• List conversations and Get conversation details - Added nullable user_message_count and agent_message_count (integer).
• Get conversation details
  • WhatsApp conversation info adds nullable call_result (answered, voicemail, busy, no_answer or failed) and opening_message (template_name, wamid, sent_at_unix, delivered_at_unix, read_at_unix, replied_at_unix, failed_at_unix, error_code, pricing_category and billable).
  • WhatsApp call permission requests add nullable responded_at_unix (integer), permanent (boolean) and granted_by_user_call (boolean).
  • Agent definition source enum adds architect.
• Update document - Added optional enable_auto_sync (boolean), auto_remove (boolean), minimum_frequency_days (integer) and tighten_only (boolean, default false).
• Update triage ticket - Added optional qa_comment (string).
• List test invocations - Added optional created_by_user_id query parameter (string).
• Create test and the other test endpoints - Tool call evaluations add nullable guardrail_trigger (object with type any).
• Rebase a branch onto main (POST /v1/convai/agents/{agent_id}/branches/{branch_id}/rebase) - The request body accepts application/json.

Speech to Text

• Realtime (/v1/speech-to-text/realtime)
  • Added optional entity_redaction (string or array of strings) and entity_redaction_mode (redacted, entity_type or enumerated_entity_type, default enumerated_entity_type) query parameters.
  • Added the redacted_transcript server message.
  • session_started config adds entity_redaction (array of strings) and entity_redaction_mode (string).

Workspaces

• Service account and workspace group permissions - PermissionType adds agents_architect, and workspace group permissions add architect.