On August 11, Mario Zechner, founder of Pi Harness, reposted data from developer 0xEvan, who used Pi to call DeepSeek V4 Flash, processed nearly 1 billion input Tokens, achieved a cache hit rate of 99.93%, and spent only 2.65 USD; without caching, the same usage was estimated to cost 132 USD. Developer Shantanu Goel stated that DeepSeek V4 Flash cache hit rates in other Harnesses usually range from 94% to 97%, but consistently exceed 99% in Pi. Mario commented, "This works especially well when using local models." In May, Mario had evaluated the Pi and DeepSeek V4 combination as "pi + ds4 == sovereign AI enterprise ready clearly." Composio, a company developing tools for AI agents, conducted a public comparative test using the same model DeepSeek V4 Flash across 8 different agent Harnesses to complete 30 highly difficult agent tasks. Pi Agent passed 20 out of 30 tasks, a success rate of 66.7%; Oh My Pi ranked second with 17 passed tasks; Claude Code, Codex, and Deep Agents each passed 16 tasks; Prime Agent and Hermes Agent each passed 15 tasks; OpenCode ranked last with 14 passed tasks. With the same model, changing only the Harness raised the success rate from 46.7% to 66.7%, a