1.13.0a3

CrewAI Inc. has pushed a significant pre-release update, version 1.13.0a3, for its open-source framework used to orchestrate autonomous AI agents. The release, led by contributor greysonlalonde, focuses on operational improvements and forward compatibility. A major new feature is the emission of token usage data within the `LLMCallCompletedEvent`, giving developers crucial visibility into the cost of running their AI workflows. This is paired with a critical bug fix that properly handles OpenAI's newer GPT-5.x models, which do not support the traditional `stop` API parameter, ensuring the framework remains compatible with the latest model families.

Beyond new features, the update includes substantial refactoring work to modernize the codebase. Core components like the `Flow` class and various `LLM` classes have been converted to use Pydantic BaseModels, a move that enhances data validation, serialization, and overall developer experience through better type hints and documentation. The release also improves project documentation, adding a comprehensive Single Sign-On (SSO) configuration guide and refining the agent capabilities overview. This release solidifies CrewAI's infrastructure as teams increasingly rely on it to build complex, multi-agent systems for tasks like automated research and content generation.