“Anthropic's use of silent performance degradation for AI research use cases makes it impossible for researchers to reliably use or plan around using Claude models for their work.”