MathCode 0.2.0 is released, introducing prefix cache shaping to significantly reduce API costs by 90% According to Beating monitoring, the Math AI team has released a mathematical form and theorem proving AI agent MathCode 0.2.0, which reduces API costs by up to 90% through prefix caching request shaping and policy control. The new version addresses the high overhead of long-range proofs and multi round interactions for large models, optimizing the stability of prompt word structures to improve prompt cache hit rates. In the scenario of theorem proving, the model needs to frequently read large code libraries and axiom libraries. MathCode 0.2.0 introduces prefix cache request shape diagnosis and policy controls mechanism. The system blocks the header attribution of prompt words to ensure structural stability, and adds a multi breakpoint cache 'MATHCODE_CHE_SMULTIBREAKPOINT', an explicit minimum prefix gateway, and an idle aware TTL strategy. By aligning the underlying request shape, the native cache hit rate of the provider is maximized, and long session bills are compressed to 10%. The new version also strengthens task management and deep thinking control. To prevent long-term overspending, the system supports setting token budget limits through MATHCODE_GOAL_MAX_TOKEN-BUDGET and limiting nested oblique slash commands to prevent infinite loops. In addition, it is demonstrated that the session supports dynamic adjustment of thinking depth through -- effort or/effort, providing levels such as low, medium, high, and max, and supports returning to default configuration at any time. In terms of compilation acceleration and deployment, the new version supports the external compiler Kimina Lean Server developed by Project Numina to run as a child process resident and route compilation check requests to the/verify interface. At the same time, the installation script 'setup. sh' has undergone a refactoring, with the addition of commands such as' status' and 'clean'. It supports release verification, automatic validation, and reuse of the Lean/Lake environment system, reducing deployment barriers. [Original link]