90-Run Benchmark Compares ThinkingCap, Fable Fusion, and Stock Qwen3.6-27B

A systematic 90-run benchmark evaluated ThinkingCap, Fable Fusion, and stock Qwen3.6-27B on agentic tasks using isolated Kubernetes workspaces and OpenTelemetry tracing. The test revealed that while ThinkingCap reduced thinking token usage by 34% and was the fastest in most tasks, the stock Qwen model remained the most disciplined and reliable overall. This empirical comparison highlights that fine-tuned LLM variants designed for reasoning do not always outperform their base models in agentic workflows. It provides concrete data showing that efficiency gains in thinking tokens can come with trade-offs in reliability and consistency. ThinkingCap's efficiency was bimodal, producing both the cheapest runs and the most expensive failures due to chasing phantom errors, while Fable Fusion proved to be an excellent investigator but prone to hallucinating facts. The benchmark utilized llama.cpp through llama-swap on a single RTX 5090 GPU with a 131k context window.