The subscription card, the window view and the capacity charts, explained on real usage. Every figure on this page is one owner's own, read from the application on October 29, 2026; accounts are letters and machines are numbers.
The subscription card
One card per subscription answers three questions: how much of the window is used, how fast, and when it resets.
- Provider, plan, source
The coloured dot is the provider, the chip the plan. “From account” means the percentage is the provider's own meter reading, not an estimate. The account is a letter (A); its real name never leaves the application.
- Used against time
100% of the limit is used while 19% of the window's time has passed. The bar fills with use; the black mark is the time. A fill ahead of the mark means the window is being used faster than an even pace.
- Pace ×
5.14× is the used share divided by the elapsed share; 1× is an even pace. The number turns ochre when use is more than 8 points ahead of time. Early in a window it swings widely, so read it together with the bar.
- Reset and freshness
“reset 5d 15h” counts down to the window's end. The dot and the age (2 min ago) tell how fresh the percentage is: a full dot is recent, a half dot stale, an empty one means no reading.
- Tokens and API-eq
3.87B tokens counted from the local history of this window, and what they would cost at public API prices: $832.90. It measures work, not a bill. It leaves out 304.64M tokens of a model with no catalogue price.
- Credits and free resets
The prepaid credit balance (58,025) converted into windows: at most about 2 windows, ≈ $2,321. Below it: free resets still available and when the nearest one expires (Oct 29).
The window view: four numbers
Open a window and its first row answers four different questions: how much is used, how long is left, where it will end, and how fast.
- Used
47% of the window is used; 53% is left, worth about $1,491.94 in API-eq. The black mark is elapsed time (60%), so the bar shows at a glance whether use is ahead of time.
- Until reset
Time left (2d 18h) and the moment of the reset (Nov 1 11:00 UTC). The grey bar is elapsed time; a window is a fixed seven days.
- At reset
Where the window ends if the current pace holds: 78%. The bar spans 0–200%; the black mark is the 100% limit and the red zone beyond it is overrun. The dotted “typical” value needs 2 earlier windows of the same plan, here 1 of 2, so it stays empty instead of guessing.
- Pace now and to last
0.78× is the pace against an even one. “Now” is how fast the limit is used (0.46 pp/h); “to last” is the speed that would make the remaining 53% last until the reset (0.80 pp/h). When “now” is longer, the limit runs out early.
The usage chart
Seven days of one window: the meter's readings, where they are heading, and what each day added.
- Cumulative line
The solid line is the percentage the provider reports, reading after reading; the pale area under it is the same course. It is a step line of real readings, not a smoothed curve.
- Now
The dotted vertical line marks the moment of the data (Oct 29 16:16 UTC). Everything to its left was measured; to its right there is only forecast.
- Projection and the ✕
The dashed brick line continues the current pace to the reset. This time it stays below the 100% line, so there is no ✕: at the right edge “78%” is where that pace would end the window.
- Account pool
The violet line is a separate pool on the account — here a dollar grant of $250 — read from the account, not from the meter, on the same percentage scale. It spent $0.00 in this window.
- Daily bars
Each bar is what the meter gained that day, in percentage points; the tallest (Oct 27) gained 16.0%. The violet cap is the pool's spend that day. A day with no data yet is an empty slot, never a zero.
- Reset marker
The vertical line on the right is the reset; ⟳ and its time (Nov 1 11:00 UTC) label it. After it the meter returns to 0% and a new window begins.
What subscription capacity is
A percentage means little until you know what it buys. Capacity is the API value of a whole window: what 100% would carry.
- Why API rates
A subscription has no price per token, so tokens are valued at the public API price of the model that produced them. That gives one yardstick for every provider; it is the value of the work, not what was paid.
- Why divide by the meter
The provider reveals only a percentage. If 63 points of the meter carried $705.53, then 100 points carry $705.53 ÷ 63 × 100. Nothing else is assumed.
- ≥ means at least
When some of the traffic has no price in the rate catalogue it is left out, so the true capacity is at least the figure shown and the figure carries ≥.
- Small samples
A cycle in which the meter rose by less than 10 points is not drawn, and medians use cycles of at least 20 points. A small rise multiplies any error in the reading.
The cycle above is a real one: Oct 27 – Oct 28, Pro · ×20.
Comparing capacity over time
Each bar is one weekly cycle of one subscription: the capacity in that cycle, the API-eq value of 100% of the 7-day window. Followed over months, the bars show whether a plan became worth more or less.
- One bar, one cycle
Width is the cycle's span, about a week and shorter when a reset came early; height is the capacity: the API-eq value of 100% of the window in that cycle.
- Colour is the plan
Green is ×1, blue ×5, violet ×20. A change of colour is a change of plan, so compare bars of one colour.
- Hatched means at least
A hatched bar rests on traffic with some unpriced calls, so it is a lower bound of the real capacity.
- ×2 is a promotion
Cycles measured under a declared promotion carry its factor above the bar; their capacity includes the bonus.
- Monthly medians
The dashed lines are the median of August 2026 and of October 2026 for Pro · ×20: $2,535 and $1,154.
- A gap is a gap
A period with no measured cycle stays empty; it is never drawn as zero.
Codex capacity fell. On Pro · ×20 capacity had a median of ≥ $2,535 in August 2026 (11 cycles); in October 2026 it is ≥ $1,154 (26 cycles), 54% less. The same direction on separate accounts: subscription A ≥ $2,535 → ≥ $1,187 (Aug → Oct); subscription D ≥ $1,401 → ≥ $1,120 (Sep → Oct). Plus · ×1 went from $148 in March to ≥ $53 in October.
Pro · ×5: the end of a promotion, not a drop. Cycles marked ×2 were measured under a declared promotion (median ≥ $1,178, May 2026); the first month without it is ≥ $553. They are left out of every trend and per-dollar figure here.
What the numbers cannot say. Capacity is API value per 100% of the meter. It falls when the provider lowers a limit, but also when the mix of models or the share of cache changes what each token is worth. The numbers alone do not tell which of these happened.
| Month | Plus · ×1 | Free | Pro · ×5 | Pro · ×20 |
|---|---|---|---|---|
| February 2026 | $922× | — | — | — |
| March 2026 | $1483× | $132× | — | — |
| April 2026 | $985× | $31× | — | — |
| May 2026 | $1291× | — | ≥ $1,1785× | — |
| June 2026 | — | — | ≥ $5533× | — |
| July 2026 | ≥ $1151× | — | ≥ $5292× | ≥ $2,07810× |
| August 2026 | ≥ $1084× | — | — | ≥ $2,53511× |
| September 2026 | ≥ $716× | — | — | ≥ $1,40113× |
| October 2026 | ≥ $5310× | — | — | ≥ $1,15426× |
Every change we observed, dated, with the plans and the number of subscriptions that show it, is on the Plan capacity page →
Capacities of different providers
The same measure for every provider and plan: the capacity of the plan in the same calendar periods, and what each dollar of it buys now. Compare plans within a column: the same weeks, the same time.
| Plan | Cycles | Now | 0–30 d | 30–60 d | 60–90 d | earlier (>90 d) | Max | Per $1 |
|---|---|---|---|---|---|---|---|---|
Anthropic / Claude Pro · ×1 · $20 / mo | 4 | $1,041Sep 23 | — | $641$228 – $1,407 · n = 4 | — | — | $1,407Sep 20 | —last measured Sep 23 |
Anthropic / Claude Max · ×5 · $100 / mo | 1 | $406Sep 23 | — | $406n = 1 · single reading | — | — | $406Sep 23 | —last measured Sep 23 |
Anthropic / Claude Max · ×20 · $200 / mo | 6 | ≥ $2,752Oct 24 | ≥ $2,596$2,232 – $2,752 · n = 5+17% | $2,222n = 1 · single reading | — | — | ≥ $2,752Oct 24 | ≥ $56.45≥ $11,289 / mo · n = 5 |
OpenAI / Codex Free | 3 | $3Apr 11 | — | — | — | $12$3 – $13 · n = 3 | $13Mar 29 | —free plan, no price |
OpenAI / Codex Plus · ×1 · $20 / mo | 31 | ≥ $52Oct 19 | ≥ $53$47 – $61 · n = 10−51% | ≥ $72$52 – $100 · n = 5 | ≥ $108$101 – $144 · n = 4 | ≥ $111$85 – $153 · n = 12 | $153Mar 27 | ≥ $11.56≥ $231 / mo · n = 10 |
OpenAI / Codex Pro · ×5 · $100 / mo | 5+5 promo, left out | ≥ $502Jul 10 | — | — | — | ≥ $553$502 – $612 · n = 5 | $612Jun 3$1,247 ×2 promo | —last measured Jul 10 |
OpenAI / Codex Pro · ×20 · $200 / mo | 60 | ≥ $1,120Oct 28 | ≥ $1,187$1,052 – $1,561 · n = 27−54% | ≥ $1,456$1,159 – $2,012 · n = 13 | ≥ $2,579$1,960 – $2,904 · n = 10 | ≥ $2,078$1,401 – $2,316 · n = 10 | ≥ $2,904Aug 13 | ≥ $25.82≥ $5,163 / mo · n = 27 |
Google / Gemini | — no weekly window readings | |||||||
xAI / Grok | — no weekly window readings | |||||||
- Now
The newest completed cycle of the plan, with its date. A reading older than 14 days is faded, so a stale reading is not set beside a fresh one.
- Same time in every column
0–30, 30–60, 60–90 and >90 days are counted back from the date of the data, for every plan. Plans were measured in different months, so compare them within a column, never along a row. n is the number of cycles; one or two are faint, because that is a reading, not a median. The small percentage is 0–30 d against the nearest older period: brick for a drop, green for a rise, grey under 5%.
- Max
The highest cycle ever measured for the plan, with its date. Cycles under a declared promotion are left out of every figure; if the highest one was a promotion cycle it is shown below, with its factor.
- Per $1 of the plan
API-eq per dollar of the monthly price, from 0–30 d only, for every plan. The monthly figure is capacity × 30.4375 ÷ 7 (about 4.35 windows); per $1 is the same in a weekly or a monthly view. The monthly figure is a ceiling: it assumes every window is used in full.
- A dash says why
Where there is nothing to compare the cell is a dash, never a zero: no cycle in that period (the plan then shows when it was last measured), a free plan without a price, or a provider without a weekly meter.
A median hides jumps: the spread of each period, a standard month with its banked and automatic resets, and the change log are on the Plan capacity page →
Other useful figures
Beside the chart the window view keeps two panels: the account's counters and the totals of this window from the local index.
- Counters
Every meter of the account at once: the 5-hour session, the 7-day week and a model sub-limit (↳ with its own denominator, not part of the 7d). The black mark is elapsed time; ⟳ is the time to that meter's reset.
- Extra usage
Whether the account may bill paid usage beyond its limit. It is disabled here.
- Tokens and cache
4.03B tokens in this window; 99.7% of them were read from cache and only 11.52M were new. A high cache share is what keeps long sessions affordable.
- Per day
The window's tokens spread over the days elapsed so far: 954.85M a day. A quick way to compare windows of different age.
- Machines
How the window's traffic splits across the machines that wrote it: machine 1 91%, machine 2 9%. Real names stay in the application.
- Calls, API-eq, grant
13,346 model calls worth $1,323.04 at API rates; the grant row is how much of the account's dollar pool was spent in the window ($0.00).