Agent Model Benchmark skill and the Yalc Framework

The Agent Model Benchmark kit answers one question. Is the cheaper model good enough for the work your agent actually does. It runs Claude Fable 5 and Kimi K3 headless on the Claude Code skills you already use, blind judges every pair of outputs with Opus, and reports quality, latency, cost, and pass or fail per task in a single results page.

Setup covers the three ways to connect Kimi to Claude Code, including the CLAUDE.md note that lets Claude Code delegate subtasks to the kimi CLI. No API key is needed beyond the claude and kimi logins you already have. In the first real 11 task run, the two models showed correctness parity on every objective check. Kimi won the lean structured tasks at about a third of the price per token, Fable won every writing and research task, and Kimi ran 2 to 5x slower with 2 errors on heavy agentic jobs.

The verdict from that run. Kimi K3 for high volume correct answer work you clean up after. Fable 5 for the polished stuff you cannot babysit. One trap worth knowing before you pipe kimi output anywhere. kimi -p quietly prepends a bullet that breaks strict JSON, and the kit strips it for you. A sample report from the real run ships in the repo so you can see the verdict format before spending anything.