DeepSeek's V4 Flash has topped model leaderboards and been hailed by developers as a "total monster" since its rollout. But in real-world testing, it completed just 53.8% of a batch of complex agent tasks. Composio ran the model through eight different agent harnesses, including Claud