Agent coding burns 1-2B tokens a day: Reddit dev does the math on a $1,000+/month API bill
JamesGooning · reddit · 2026-11-13
A developer sharing token-consumption math for agent coding: running local QFN as a daily driver, his logs show 1-2 billion tokens burned per day — and the GPU isn't even saturated, since the LLM mostly waits on shell/bash. Theoretical max is 950M uncached tokens/day (11,000 × 3,600 × 24).
- On commercial APIs (Haiku/Luna-class, 90% cache hits) that's roughly $10/day or $300/month for a single serial project
- Real productivity means multiple parallel projects, plus his wife now shares the harness for searches — a purely API-based setup would exceed $1,000/month
- Local isn't cheap either: a single RTX 6000 is the only GPU that can serve the household, and always-on agent load may 4x soon
He asks how others keep API-based agent coding affordable long term.
More from coding & agent
- "Code review is dead": nobody actually reads an 8,294-line agent diff — prasenx · 2026-11-13
- AI educator alfcnz: students lack coding-agent skills, so he'll teach a method — alfcnz · 2026-11-13
- Decision models for agent tool selection: 50 tools to 3, 47% fewer tokens, 22% faster — rajistics · 2026-11-13
- Anthropic engineer: Claude iterated meeting calendar designs live in one brainstorm — Flomerboy · 2026-11-13
- Nuance Live preview: build real-time audio-video apps where an AI persona calls your tools — koltregaskes · 2026-11-13
- Agent Deleted 6 Lines, Caught 3 Weeks Later: 400 Customers Wrongly Charged — Imaginary-Mix-7777 · 2026-11-13