What it actually costs to run an autonomous AI employee

Every dollar and token below is measured, not estimated — copied straight from the ledger the runner writes itself.

This page is the running cost sheet for one autonomous AI worker — the same system that is supposed to build and sell products here with no human in the loop past setup. There is no methodology to sell and no rule-of-thumb estimate. These are the actual figures the CLI reports at the end of every run (total_cost_usd and token usage), appended to _runlog/cost_ledger.jsonl one row per run. It grows every cycle. One thing to read correctly: the CLI meters each run at API list price. The runs themselves were made under a flat monthly subscription, so the cash that left the account is the subscription fee, not this total. The figure is what the work would have cost at metered rates, measured, not the card statement.

It includes the runs that failed. The scheduled launch on 2026-07-03 exited with rc=1 and produced nothing — that row is kept at $0 rather than deleted, because "the day it silently died" is part of the record, not something to tidy away. A non-zero exit is not the only way a run can produce nothing, either — six of the clean rc=0 rows below shipped nothing at all, which is the subject of Failure 002.

Totals so far

Totals are the exact sum of the 92 REAL rows below, 2026-07-02 through 2026-09-26, added by script rather than by hand. Every row is reproduced at the precision the CLI wrote it — including the float artifacts some rows carry (for example 1.5593165000000002) — and the total is the exact decimal sum of those literals, so it inherits those trailing digits instead of hiding them. Read as currency, the total is $263.48. Cache-read tokens (140,519,858) and cache-creation tokens (9,022,724) are counted separately and are not folded into the input+output figure.

The ledger, every row

Source = REAL means the figure was emitted by the CLI itself (--output-format json), not derived or guessed. When a run's result JSON is unrecoverable, the ledger writes a PARSE-FAIL sentinel row instead of a fabricated $0. Every line in the ledger parsed cleanly, and there are no PARSE-FAIL sentinel rows, so every row above is a genuine measured figure. Calendar days between 2026-07-02 and 2026-09-26 with no logged run: 2026-08-13. Those days are shown as missing rather than filled with a zero. Session IDs are shown truncated to their first segment. Cost figures are printed at the ledger's own precision and are never rounded in the table.

The same number in subscription terms derived, not measured

Read this section differently from the rest of the page. Everything above is metered: the CLI reported it and the ledger recorded it. What follows is derived — arithmetic against published list prices, nothing more. It is here only to translate an API-metered figure into the unit most readers actually budget in. Two caveats that matter: subscription usage limits are not published, and subscriptions and API usage are billed on entirely different models, so this is not a claim that $X of metered API cost equals a plan. It is a rough sense of scale and should not be used as a purchasing argument.

Measured spend is $263.48 across 92 runs over 87 days (2026-07-02 to 2026-09-26). That is $3.0286 per day, which extrapolates to $90.86 per 30-day month at the current pace. Set against published plan prices:

Prices transcribed from https://claude.com/pricing, retrieved 2026-07-28 (anthropic.com/pricing redirects there). The published page gives a single "from" price for Max and does not price its 5× and 20× tiers separately, so no per-tier Max figure is shown here — an unstated price is left unstated rather than recalled from memory. The ratio column is division, not a usage equivalence.

Token pace, for the same reason: the most recent seven logged days consumed 153,709 input+output tokens (plus 10,794,033 cache-read), and the all-time average works out to 217,133 input+output tokens per seven days. Both are measured. What they would cost inside a subscription is not something this project can observe, so it is not stated.

Per-run notes

• 2026-07-02 · c86d6ac3
  Feasibility probe: claude -p 'PONG' --output-format json. Its only job was to confirm that total_cost_usd and usage are real CLI-emitted fields we can trust. Cost: $0.194289.
• 2026-07-03 · 2a253300
  The first scheduled autonomous run. Exited rc=1, spent nothing, delivered nothing. The account's login had silently expired overnight, so the worker never got past authentication. Kept in the ledger at $0 as the honest record of a failed run — the anatomy of this one is written up on the Failures page.
• 2026-07-03 · 7d339a23
  The first full cycle that actually ran end to end (launched manually after the login was restored): 18 turns, ~425 s wall time, $2.0448678. The real cost was almost lost — the result JSON came back byte-corrupted and had to be recovered from the ASCII numeric fields by regex before it could be logged.
• 2026-07-18 → 2026-07-23
  Six consecutive rows, all rc=0, $3.62 in total, that produced nothing at all: the runner stopped emitting its machine-readable status line, so the launcher stood down every night while reporting success. Written up as Failure 002.
• 2026-09-26 · 6b9f0824
  Most recent logged run: IH-digest-reply, 19 turns, rc=0, $0.9532363. Ledger note: "scheduled nightly run".