Llm

Caveman is a Claude plugin that asks the model to answer in terse fragments. Does it save money? Sometimes—mostly when the model would otherwise produce a very long answer. I compared Caveman’s lite, full, ultra, and wenyan modes with two baselines: no added instruction and “Answer concisely.” The benchmark covered 15 prompts, three runs per condition, and two Claude Opus models. I measured token cost, not answer quality. Model and task Caveman cost vs. concise Claude Opus 4.7, short Lite +10%; Full +3%; Ultra +3%; Wenyan +10% Claude Opus 4.7, long Lite +9%; Full +10%; Ultra −5%; Wenyan −1% Claude Opus 4.6, short All modes: 3–16% less; uncertain Claude Opus 4.6, long Lite −54%; Full −56%; Ultra −59%; Wenyan −58% Positive numbers mean higher cost; negative numbers mean savings. The practical rule: Caveman has to save enough output tokens to pay for its added instructions. That is easy with a long tutorial and hard with a short answer. ...

TL;DR: For a small, well-designed API (10 tools), structured MCP tool calls consistently outscore code generation on correctness — 0.99 vs 0.97 — and the gap concentrates in tasks where domain-specific logic matters. Adding a reference document to MCP tools costs 6–14% more tokens with zero accuracy gain. Cloudflare’s search+execute pattern matches MCP accuracy but uses more tokens. With MCP tools, Haiku is within 1% of Opus at 1/12 the cost. ...

Distillation reveals which capabilities are economically reproducible.