We’ve released Conversational AI Engine v2.7 on May 20, 2026. This release expands our MLLM and avatar provider ecosystem, improves greeting control, and introduces important updates to API status codes, conversation turn retrieval, and custom instruction behavior.
What’s new in v2.7
New MLLM provider: xAI Grok
- Expands our multimodal model options, giving customers more flexibility when building real-time conversational AI experiences.
New Avatar provider: Generic avatar (Beta)
- Adds a new avatar integration option, further broadening our avatar ecosystem for multimodal and visual agent use cases.
Greeting interruption control
- A new interruptable parameter is now supported under properties.llm.greeting_configs.
- This allows clients to control whether user speech can interrupt the agent greeting during playback, enabling more predictable first-turn behavior.
Improvements you’ll notice
More granular API error handling
- RESTful API response status codes and reason values have been updated. New status codes include 401, 429, and 500, and new reason values include ServiceNotEnabled, AccountSuspended, and ResourceAllocationFailed.
- The previous InvalidRequest reason has also been replaced by more specific values: InvalidRequestBody, MissingRequiredField, and InvalidFieldValue.
- If your integration relies on reason values for error handling, retry logic, or alerting, please review and update your logic accordingly.
Paginated conversation turn retrieval
- The Query conversation turns API now returns paginated results instead of all turns in a single response. This improves scalability for longer sessions, especially those with more than 50 conversation turns.
- The response now includes additional fields such as agent_id, name, channel, total_turn_count, and pagination.
New post-session turn notification event
- A new 112 turns finished notification event is available. It delivers conversation turn data in batches after a session ends, providing another option for post-session analysis and processing.
Updated custom instruction behavior
- The default value of on_listening_action in the Send a custom instruction API has changed from inject to interrupt.
- When set to interrupt, the API immediately interrupts the current flow and starts a new round of dialogue. If your integration depends on the previous behavior, please explicitly set on_listening_action to inject
For full technical details, please refer to the release notes