DeepSeek harness burns 12.5x more tokens than Pi — 29 built-in plugins and 70% tool definitions are why
teortaxesTex · x · 2026-09-25
A teardown of Pi vs the new DeepSeek harness (GUI) shows massive token overhead: across three test tasks the harness used 69,418 tokens vs Pi's 5,532 — a 12.5x gap — with identical results and Pi even faster.
Context breakdown:
- Tool definitions are the biggest culprit: 70% of the harness's single-request context (70–72% text tasks, 62% code tasks), because standard mode ships 29 built-in plugins; Pi disables tools on text tasks and keeps just 4 CRUD tools for code
- System prompts: 19–20% (code 17%) for the harness
- User messages, history, and tool returns: 8–10% (text) / 21% (code)
System prompts and user messages are similar between the two — the token gap mostly comes from plugin tool definitions in standard mode; minimal mode should cut consumption substantially. The author plans further analysis.
More from coding & agent
- Micius-Flash: 27B Multimodal Model for Hardware and Semiconductor Agents — zhiman-ai · 2026-11-11
- Jane Street's AI Skill Is a Strict Socratic Tutor You Can Copy Today — omojumiller · 2026-11-11
- Designer builds 100k-site German heritage map with Claude, zero code written by hand — neutronenkind · 2026-11-11
- Matt Pocock: DDD, not agentic tricks, is what blows minds in his AI course — mattpocockuk · 2026-11-11
- AI agents run 24/7 for a month to rewrite Météo-France's AROME weather model — IgorCarron · 2026-11-11
- Prompting secret: ask the model to roleplay a legendary high-agency explorer — tokenbender · 2026-11-11