DeepSeek's top-ranked V4 Flash stumbles on real agent tasks as its prices surge

DeepSeek V4 Flash has topped AI leaderboards, yet it struggles with real-world agent tasks. Testing by Composio across harnesses like Claude Code and OpenCode showed the model completed only 53.8% of complex, multi-step workflows involving Gmail, GitHub, and Slack. Out of 240 runs, only 129 passed, with only six workflows succeeding across all tests. These results suggest that orchestration and reliability are more critical for enterprise success than raw model capability or low pricing.