SubWorthDay 30/14 · measuring

The experiment, as it runs

A passive collector records the quota reading behind every status line refresh during ordinary work. This page is the aggregate of that log, regenerated and republished daily by a local scheduled job — no server, no live endpoint, no number that has been smoothed on the way here.

Plans at a glance

Two subscriptions, one method

Each plan is measured on its own quota signal and priced against its own API rate card. The figure is the latest rolling estimate of what one weekly window is worth; the detail for each plan is under its tab.

Claude Max 20x$200/month
$1,945
per weekly window · n = 54 · since 2026-08-30 02:18 UTC
Codex Pro 20x$200/month
$1,214
per weekly window · n = 58 · since 2026-09-05 14:29 UTC

Day 30 of 14

Claude Max 20x

$200/month

390,714
Quota snapshots
629.7 h
Collection span
133
Sessions observed
620 /h
Snapshot rate
2
Quota windows

Started 2026-08-30 02:18 UTC · latest reading 2026-09-25 08:00 UTC · Claude Code 2.1.247, 2.1.251, 2.1.252, 2.1.257, 2.1.258, 2.1.259, 2.1.260, 2.1.261, 2.1.263, 2.1.266, 2.1.267, 2.1.268, 2.1.269, 2.1.270, 2.1.273, 2.1.274, 2.1.278, 2.1.280, 2.1.281 · Fable 5, Fable 5.1, Opus 4.8, Opus 5, Opus 5 (1M context), Opus 5.5 (1M context), Sonnet 5

Quota windows

Percentage used over time

Readings are downsampled to one point per UTC hour. Both axes are fixed: 0–100% of the window, so a small move looks small.

5-hour window

0% used252,646 snapshots · 276 hourly points
0255075100%Aug 30, 02:00Sep 24, 21:00

7-day window

14% used390,714 snapshots · 390 hourly points
0255075100%Aug 30, 02:00Sep 25, 08:00
WindowUsedObserved rangeDistinct valuesSmallest stepSession spreadResets seenWindow ends
5-hour0%0–55%561 pp21 pp692026-09-24 21:40 UTC
7-day14%0–75%761 pp22 pp42026-09-29 05:00 UTC

“Session spread” is the widest disagreement seen between two sessions reading the same window at the same moment. “pp” = percentage points.

Rolling estimate

What one window of quota is worth, so far

Estimated

Each point is the API-equivalent value of one whole quota window, computed from every high-quality observation recorded up to the end of that UTC day — the day-6 point restates the estimate on six days of evidence rather than reporting day 6 on its own. The shaded band is ±1 standard deviation of that sample. A band that narrows against the line is the convergence this experiment is looking for; the line itself moving is not.

5-hour window

$358 per windown = 376 · through Sep 25
$0$250$500$750$1,000n=376Aug 30Sep 25

7-day window

$1,945 per windown = 54 · through Sep 25
$0$1,000$2,000$3,000$4,000n=54Aug 30Sep 25

One observation is a stretch of a single session in which the window moved at least 3 percentage points without crossing a reset, priced from the local token log against the published API rate card. n counts the observations a point rests on.

These are estimates, not results. The method is still inside its 14-day validation window, and the day-14 criteria can void every number in this section. Nothing here is an official quota, and no Worth Multiple is published until those criteria pass.

Day 0

What the first readings already changed

The field this benchmark depends on is specified nowhere. The first readings measured four properties, and each one moved a rule in the methodology.

The percentage is quantized to whole numbers

Every reading lands on an integer, so a one-point delta carries up to half a point of rounding error — hence prefer few large deltas, and down-weight anything near the quantization floor.

evidence: integer-quantized in every window · 5-hour 56 distinct values · 7-day 76 distinct values

Each session serves its own cached reading

Two sessions on one account can report different percentages at the same second — so deltas are computed inside a single session’s series, never interleaved across sessions.

evidence: widest concurrent disagreement 22 pp

The reset timestamp is the end of the window

Second-precision, hour-aligned, always in the future — a reliable boundary marker, so an interval straddling a reset is detected and discarded instead of silently producing a negative delta.

evidence: hour-aligned not confirmed · 73 window resets crossed so far

Passive collection is cheap in effort, not in bytes

At a five-second refresh across concurrent sessions the log grows by roughly 620 mostly-identical records an hour — the shipping collector will collapse them; the experiment keeps them so the sampling cadence stays measurable.

evidence: 390,714 snapshots over 629.7 hours

Provenance

How this page is produced

~/.subworth/snapshots.jsonl          local, append-only, never uploaded
        ↓  bun bin/export-site.ts    aggregate + de-identify
site/src/data/live.json              hourly curves + counts only
        ↓  git push                  static build, no backend
this page

The exporter builds its output field by field from literals; nothing is copied out of a snapshot. Session identifiers become a count, and no path, prompt, or model output exists in the log to begin with. The published file is the whole of what this site knows — the field list is here — and it ships in the repository as site/src/data/live.json.

Updates run once a day from a local scheduled job. The timestamp in the footer is when the export ran, not when you loaded the page — this site has no live connection to anything.

Not here yet

What is deliberately missing