I priced every Grok Bot call: $20 to $200 plan in 5 days, long jobs eat 52%
jonathanmalkin · reddit · 2026-10-16
The author's Grok Bot usage spiraled: free tier exhausted same day, $20 plan lasted 3 days, $60 lasted 2, a $20 top-up burned in 90 minutes, and within five days he was on the $200 plan. He spent a weekend pricing every call in the usage export at list API rates.
Where the money went
- Long jobs: 117 runs of 16+ steps cost $323, 52% of the week. Multi-step tasks re-read context every step, and history compounds.
- The front door: the chief-of-staff bot receives every request and carries the longest history—$236 in one week, double any other bot; its priciest single job ran 222 steps for $15.31.
- Cold reloads: 248 calls loaded 50k+ tokens with nothing cached, adding $69.
- Small asks: five one-line voice edits to one email cost $1.26, mostly re-reading a 120k-token chat.
- Scheduled jobs—the obvious suspect—were only 8%.
Why: 90% of tokens were re-reads of context the bot already had. History length and step count drive cost; a blank bot loads 35k tokens on first reply, the chief-of-staff bot carries 169k. All chats on one bot share a single context, so new chats don't reset it. Cached-token pricing also varies: Grok 4.7 charges a quarter of input price, Claude Opus 5.5 a twentieth.
Why the bot can't self-limit: no step-limit setting in docs; staff admit chat instructions are "only a hint they can ignore"—a written handoff rule was broken 117 times; no tool shows a bot its own spend; "a limit is not a hard stop mid-run."
Advice: measure first by pricing the usage export (the author's guess was wrong); script what repeats (no conversation to re-read); hand long jobs to a clean-context agent (cloud agent, Codex, Claude Code) before they start and read results from a file. Note: analysis predates the unified X plan, and Grok Bot sharing quota with Cursor may nullify the handoff trick.
More from coding & agent
- Dev uses Grok to clean up 3D printer firmware, catches it sneaking in nonsense code — yacineMTB · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16
- Yacine: vibe-coded repos with 3,330 agent-written tests are all useless — yacineMTB · 2026-10-16
- Dev argues coding models now beat nearly all engineers in every branch of software — zack_overflow · 2026-10-16
- Q4-quantized coder model on 8-core CPU builds interactive solar system demo in one shot — IntrepidMindExplorer · 2026-10-16