DeepSeek v4 Flash Outperforms Competitors in Speed and Cost on Agentic Tasks

An empirical evaluation of three prominent open-weight models—DeepSeek v4 flash, GLM 5.2, and Kimi K3—revealed that DeepSeek v4 flash is significantly faster and more cost-effective on complex, multi-application agentic workflows. Despite consuming more tokens on average, DeepSeek completed tasks in 164 seconds at a cost of just $0.08 per task, while maintaining a success rate comparable to its competitors. This comparison highlights that lightweight, fast, and highly cost-effective models can perform agentic workflows at near-frontier model success rates for a fraction of the cost. It signals a shift where developers can build highly efficient AI agents using open-weight models without sacrificing performance. The evaluation used Pi Agent as the harness combined with Composio MCP to access external applications like Slack, Gmail, and HubSpot, requiring all deterministic checks to pass for a successful run. While DeepSeek v4 flash had the lowest cost ($0.08) and fastest time (164s), it was the most token-hungry model, consuming an average of 569k tokens per task compared to GLM 5.2's 409k tokens.