~ / guides / context-tax

Context tax: what your setup costs on every turn

Data: 7204 public rigs · updated 2026-10-09

"My Max plan runs out in two hours" is the most common complaint about coding agents this year. A big, invisible driver is the stuff loaded into every request: instruction files, MCP tool schemas, and skill/subagent descriptions. We call that the context tax.

2.3kmedian rig (tokens)
5.0kaverage
11.5k90th percentile

Distribution

0–1.0k
2094
1.0k–2.0k
1251
2.0k–4.0k
1571
4.0k–8.0k
1163
8.0k–16.0k
670
16.0k–32.0k
307
32.0k–∞
148

What drives it

How we estimate

Characters ÷ 4 for root instruction files (plus one level of @imports) and for skill/subagent frontmatter descriptions; MCP schema sizes from a table of typical tool counts per server (2.5k tokens for unknown servers). It's an estimate for comparison, not a bill. Full method.

Leanest high-power rigs

8+ components, lowest context tax.

copied ✓