Your coding agent burns up to 12k tokens before you type: 30 harnesses span 50x
yb2698 · x · 2026-10-28
Hugging Face researcher nathanhabib1011 measured the default system prompts of 30 open coding harnesses: they range from 235 tokens (goose) to 11,974 tokens (Gemini CLI) — a 50x spread before any tool definitions are added.
Key data points: CUGA (IBM) 8,325, OpenHands 7,503, Deep Agents (LangChain) 5,843, Copilot (VS Code) 5,622, Codex CLI 4,365, Aider 3,777, Cline 1,282, mini-SWE-agent 800. This "context tax" directly affects usable context window, cost, and response quality. Measured with the o200k tokenizer via Harness Lens, prompts pinned to commits.
More from coding & agent
- Anthropic models burn 2-3x tokens and cost 10-50x more on FrontierCode 2.0 — teortaxesTex · 2026-10-28
- Devs praise OpenAI's Decisions API: Codex integration drives full-time switch — craigsdennis · 2026-10-28
- Personal AI agent trains a chess transformer on dual RTX 4090 and battles Sunfish — andreisavu · 2026-10-28
- FrontierCode 2: Anthropic models burn 2-3x tokens, cost 10-50x more than open rivals — bookwormengr · 2026-10-28
- Opus 5.5 ultracode workflow with 9 subagents builds a Three.js scene — nptacek · 2026-10-28
- Agent sandboxes talk at PLATFORMday SF: the missing runtime for AI — gethackteam · 2026-10-28