Kimi K3 completes a real coding task via OpenCode agent for $0.19

A developer tested Moonshot AI's Kimi K3 model on a practical coding task using OpenCode v1.18.3, measuring actual end-to-end cost rather than per-token pricing. The agent was given a small repository with a broken leap-year implementation and tasked with identifying the bug, fixing the source file, and passing all tests without modifying the test file. The run completed successfully in 84.3 seconds across six model steps and seven tool calls, consuming 28,707 total tokens at a measured cost of $0.19 via third-party aggregator Vancine. The author emphasizes that this single controlled task is not a general benchmark and does not guarantee similar costs or outcomes across different repositories or agent configurations. The experiment highlights that per-token price alone can be misleading for coding agents, where full-task cost depends on file reads, tool calls, reasoning, and retries.