Vapi Release Notes

What's New: Week of July 13, 2026

Playback Speed for Recordings: Call recording players now include a playback-speed control.

Download Every Recording Type: You can now download every recording a call produced: mono, stereo, separate assistant and customer tracks, video, and packet capture.

VAD Transitions in Call Logs: The call log now surfaces voice-activity-detection transitions with a per-phase latency breakdown.

HIPAA Compliance: xAI is now HIPAA-compliant across its model, voice, and transcriber.

What's New: Week of June 29, 2026

OpenAI Realtime v2: OpenAI's latest Realtime v2 model is now available for assistants.

AI-Generated Tool Failure and Completion Messages: You can now set role: 'system' on request-failed tool messages, and the dashboard has a new UI for configuring AI-generated messages when tools fail or complete. This gives assistants more natural responses when tools hit errors.

Call Logs Improvements:

• Significant latency improvements across both the dashboard and the /calls API endpoints.
• The call detail flyout now shows which assistant or squad handled the call, includes a link to the assistant or squad, and indicates which phone number was used.
• In squad or handoff calls, transcript messages now show which assistant said what, making it easier to trace conversation flow.

MCP Child Tools in Dashboard: When you connect an MCP server, the dashboard tool form now lists all child tools it discovers, so you can see exactly what capabilities your MCP server exposes.

What's New: Week of June 15, 2026

Claude Haiku (Global): Now available as a model through Amazon Bedrock for assistants.

Pronunciation Dictionary Management in Voice Config: Pronunciation dictionaries configured via the API can now be viewed and managed directly in an assistant's voice settings in the dashboard.

Stale dictionary references are flagged when a voice changes to a model that cannot apply them.

Rotating Tool Messages: You can now configure multiple message variants for a tool, and the assistant picks one at random so longer calls feel less repetitive.

Dynamic Variables in Test Calls: A dialog lets you set dynamic variable values before starting a test call from the dashboard.

Concurrency and Rate Limits in Organization Settings: Your call concurrency cap and API request rate limit now appear as read-only fields in Organization Settings.

Fixes and Improvements:

Cartesia voice overrides in squads now apply correctly instead of falling back to a hard-coded default.

Duplicate tools sharing the same function.name are de-duplicated during model streaming, preventing duplicate tool calls.

The call concurrency chart in analytics now renders correctly.

What's New: Week of June 1, 2026

xAI Speech-to-Text and Text-to-Speech: xAI is now available as a transcriber (STT) and voice (TTS) provider for assistants.

Upgraded Vapi Voices: A new text-to-speech model powering Vapi Voices makes them sound more authentic, human, and consistent — at ~50% lower cost.

Existing deployments don't change automatically — opt in by setting version: 2 on the voice configuration via the API or Dashboard. See Vapi Voices for supported voices and audio samples.

Pronunciation Dictionaries in the Dashboard: The Assistants view now supports creating new pronunciation dictionaries directly from the dashboard.

Phone Number Fixes: Improvements to phone number creation and listing.

The Phone Numbers list no longer breaks on rows with no number or SIP URI.

Creating a phone number now validates its Vapi identifier up front.

What's New: Week of April 20, 2026

Logs UX Refresh: New filter layout plus a round of UX improvements — improved date picker, active row is clearly highlighted across all log views when the flyout is opened, log tables are fully keyboard-accessible, sortable cost and duration columns, pagination, and more.

Squads context

Engineering

Plan Handoff Type — previousAssistantMessages

Forwards only the conversation history from before the current assistant's session. The current assistant's own messages and tool calls are excluded entirely from the handoff payload. See the updated handoff context configuration docs.

assistant.speechStarted Event — Live Captions & Word-Level Timing (GA)

A new opt-in message fires as the assistant begins speaking each segment, carrying the full turn text, turn, source (model / force-say / custom-voice), and optional timing:

• Per-word alignment on ElevenLabs
• Cursor-based word-progress on Minimax (set voice.subtitleType: "word", with correct CJK handling)
• Text-only fallback on all other providers

Subscribe by adding "assistant.speechStarted" to your assistant's clientMessages and/or serverMessages — now GA with no feature flag. Use it for live captions, karaoke-style highlighting, or any UI that needs to stay in sync with assistant audio. Fully backward-compatible; no existing messages changed.

Autofallbacks on Transcribers

Let Vapi pick the best transcriber to fall back to if your primary one fails — even mid-call. Opt in by setting assistant.transcriber.fallbackPlan.autoFallback.enabled to true. See the updated transcriber fallback plan docs.

What's New: October 2025 – March 2026

Here's a summary of major items shipped from October 2025 through March 2026.

Platform

• Squads v2: Visual builder to simplify sophisticated multi-assistant orchestration with seamless handoffs between specialized agents.
• Composer (Alpha): Intelligent assistant inside the dashboard that allows you to describe what you need through plain text prompts to help build, adjust, and debug voice agents.
• Simulations (Alpha): Voice agent testing feature to build confidence through enabling systematic, AI-powered testing in specific scenarios with evaluation of outcomes.
• Monitoring & Issues: Automated call quality monitoring with trigger-based issue detection, alerting, and resolution suggestions.
• HIPAA with Data Retention: New compliance mode with private storage and in-dashboard toggle/purchase flow — available for additional cost.
• Zero Data Retention: Compliance mode that keeps context data during call as needed to execute tasks and retains no data afterwards.
• Consolidated Logs: Unified log viewing into a single page.
• Vapi Voices: 12 new ultra-realistic voices released, optimized for latency and cost with adjustable speed controls exposed. 8 legacy voices deprecated.

New Models & Provider Support

Transcriber Models (Speech-to-Text)

• Deepgram Nova-3 Languages: Added Hebrew, Urdu, Tagalog, and Arabic bilingual support.
• Cartesia Transcriber: ink-whisper.
• Soniox: stt-rt-v4.

Intelligence Models (LLM)

• GPT-5 Family: OpenAI's latest intelligence models, including GPT-5, 5-Mini, 5-Nano, 5.1, 5.2, 5.4, 5.4-Mini, 5.4-Nano.
• Claude 4.5–4.6: Anthropic's latest intelligence models Sonnet 4.5, Opus 4.5, Opus 4.6, Sonnet 4.6.
• Gemini 3 Flash: Google's latest intelligence models.
• Grok 4 Fast: Reasoning and non-reasoning variants.
• GPT Realtime Mini: OpenAI's lightweight realtime model.

Voice Models (Text-to-Speech)

• Cartesia: sonic-3, sonic-3-2026-01-12, sonic-3-2025-10-27.
• WellSaid: Caruso (new), legacy.
• Inworld: inworld-tts-1 (REST, original), inworld-tts-1.5-max (WebSocket, $10/M chars), inworld-tts-1.5-mini (WebSocket, $5/M chars).
• ElevenLabs Scribe v2: Latest version of ElevenLabs speech-to-text.

Developer Tools & API

• Structured Outputs Improvements: Updates to our AI-powered analysis and data extraction tool, including transient structured outputs, audio-based extraction, and regex extraction.
• SIP Request Tool + DTMF over SIP INFO: Send SIP requests and DTMF tones via SIP INFO messages during calls.
• Variable Passing Between Tool Calls: Pass output variables from one tool call as input to subsequent tool calls.
• Encrypted Tool Arguments: Encrypt sensitive tool arguments to protect data in transit.
• Low Confidence Speech Hook: Hook that triggers when the transcriber returns low-confidence speech results.
• Time Elapsed Hook: Hook that triggers at specified time intervals during a call.
• assistant.speechStarted Event: New event fired when the assistant begins speaking.
• MCP Improvements: Bearer auth, $ref dereferencing, child tool messages/discovery.
• Warm Transfer Improvements: SIP support, caller ID, context engineering, variable filling.

Breaking Changes & API Cleanup

Legacy Endpoint Removal: The following deprecated endpoints have been removed as part of our API modernization effort:

• /logs - Use call artifacts and monitoring instead
• /workflow/{id} - Access workflows through the main workflow endpoints
• /test-suite and related paths - Replaced by the new evaluation system
• /knowledge-base and related paths - Integrated into model configurations

Knowledge Base Architecture Change: The knowledgeBaseId property has been removed from all model configurations. This affects:

• XaiModel, GroqModel, GoogleModel
• OpenAIModel, AnthropicModel, CustomLLMModel
• All other model provider configurations

Transcriber Property Deprecation: AssemblyAITranscriber.wordFinalizationMaxWaitTime and FallbackAssemblyAITranscriber.wordFinalizationMaxWaitTime are now deprecated:

• Use smart endpointing plans for better speech timing control
• More precise conversation flow management
• Enhanced end-of-turn detection capabilities

Schema Path Cleanup: Removed numerous unused schema paths from model configurations to simplify the API structure and improve performance. This cleanup affects internal schema references but doesn't impact your existing integrations.

New v2 API: We are introducing a new API version v2. These changes are part of our ongoing effort to:

• Simplify the API structure for better developer experience
• Remove redundant and deprecated functionality
• Complete the transition to new evaluation and compliance systems
• Improve API performance and maintainability

For details on the new features that replace these deprecated endpoints, see our recent changelog entries:

• Enhanced Authentication & Custom Credentials (Aug 30)
• Recording Consent & Compliance Management (Sep 2)
• Evaluation System Foundation (Sep 5)
• Evaluation Execution & Results Processing (Sep 28)

Migration Guide

Removed Endpoints

The following endpoints are no longer available:

• GET /logs - Use call artifacts instead
• GET /workflow/{id} - Use main workflow endpoints
• GET /test-suite, POST /test-suite - Use evaluation endpoints
• GET /test-suite/{id}, PUT /test-suite/{id}, DELETE /test-suite/{id} - Use evaluation management
• POST /test-suite/{testSuiteId}/run - Use evaluation runs
• GET /knowledge-base, POST /knowledge-base - Integrated into model configurations
• All related nested endpoints and operations

See Also:

• Authentication System Updates (Aug 30) - For credential management migration
• Recording Consent Features (Sep 2) - For compliance system details
• Enhanced Transcription (Sep 8) - For AssemblyAI timing alternatives

Evaluation Execution & Results Processing

Evaluation Execution Engine: Run comprehensive assistant evaluations with EvalRun and CreateEvalRunDTO. Execute your mock conversations against live assistants and squads to validate performance and behavior in controlled environments.

Multiple Evaluation Models: Choose from various AI models for LLM-as-a-judge evaluation:

• EvalOpenAIModel: GPT models including GPT-4.1, o1-mini, o3, and regional variants
• EvalAnthropicModel: Claude models with optional thinking features for complex evaluations
• EvalGoogleModel: Gemini models from 1.0 Pro to 2.5 Pro for diverse evaluation needs
• EvalGroqModel: High-speed inference models including Llama and custom options
• EvalCustomModel: Your own evaluation models with custom endpoints

Evaluation Results: Comprehensive result tracking with EvalRunResult:

• status: Pass/fail evaluation outcomes
• messages: Complete conversation transcript from the evaluation
• startedAt and endedAt: Precise timing information for performance analysis

Target Flexibility: Run evaluations against different targets:

• EvalRunTargetAssistant: Test individual assistants with optional overrides
• EvalRunTargetSquad: Evaluate entire squad performance and coordination

Evaluation Status Tracking: Monitor evaluation progress with detailed status information:

• running: Evaluation in progress
• ended: Evaluation completed
• queued: Evaluation waiting to start
• Detailed endedReason including success, error, timeout, and cancellation states

Judge Configuration: Optimize evaluation accuracy with model-specific settings:

• maxTokens: Recommended 50-10000 tokens (1 token for simple pass/fail responses)
• temperature: 0-0.3 recommended for LLM-as-a-judge to reduce hallucinations

Evaluation Capabilities