[Claude Code Runs Kimi K3: Most Expensive and Slowest, Costs Nearly 4 Times Hermes Agent] According to monitoring by Beating, AI Agent infrastructure company Composio expanded testing by integrating the same Kimi K3 model into six different Agent frameworks to complete 26 identical tasks. This comparison focuses on the frameworks, with the underlying model consistently being Kimi K3. Kimi Code ranked first, completing 21 tasks. Hermes completed 20 tasks, while Pi Agent and Claude Code each completed 19 tasks. OpenCode completed 18 tasks, and Codex ranked last with 17 tasks. Based on Kimi K3's official pricing estimates, Hermes and Pi Agent had the lowest average cost per task, at $0.39 and $0.40 respectively. Claude Code reached $1.47 per task, approximately 3.8 times the cost of Hermes. Pi Agent was the fastest, with a median time of 161.7 seconds per task. Claude Code was the slowest, taking 347.6 seconds. Using the same model but different frameworks resulted in a maximum difference of only 4 successful tasks, yet the average cost varied by nearly 4 times, and the time required differed by more than 2 times. [Original Link]