Quota measurements and evidence

Measure all available windows in the same run

Future runs record every readable allowance window using one token ledger, estimating five-hour, daily, weekly and monthly limits separately. Short and weekly windows take priority for precision; longer windows retain rough estimates with uncertainty.

Recheck meters after each batch and split intervals at resets. Mark unavailable or insufficient signals explicitly; do not scale limits by elapsed time. Preserve the model and input, output and cache mix; rounding bounds do not express overall accuracy. Publish data, reports and plan-card results together.

Measured quota data · 2026-10

Observations and conditional estimates, not official allowances. Native windows only; no automatic monthly projections.

ChatGPT Pro 200 (Codex) · gpt-6-astra

Conditional estimate · low confidence · 2026-10-28

Native window: weekly · 28 calls · Used 89% → 92%

ChatGPT Pro 200 (Codex) token usage
Source	Fresh input	Cache read	Output
Bench	2,320,149	271,104	4,070

Cache writes were zero; write weights are uncalibrated. Output includes reasoning without double-counting.

Conditional range at the observed mix: 64,883,075–129,766,150 tokens / weekly

Applies only to ChatGPT Pro 200; tier confirmed by the account owner. Astra main batch: 28 calls with unchanged controller token counters. Bounds cover meter rounding only, assuming linear consumption, settled meters and no other-device traffic. No 5-hour/monthly window returned; Sol capacity withheld due to inconsistent readings.

Measurement process and calculations (Chinese)Meter readings and per-call usage

Astra / Sol retest · Pro 200

Model / successful calls	Measured tokens	Weekly capacity (conditional)
gpt-6-astra32 calls	2,970,893	64,883,075–129,766,15028-call main batch only
gpt-6.1-sol72 calls	6,679,074	Pending: inconsistent readings; estimate withheld

Only a weekly window was returned; 5-hour/monthly capacity is unavailable. Sol tokens were measured, but final readings changed 92% → 94% → 0% with a changed reset time and controller activity. Shared pool; capacities cannot be added.

Retest methodology and anomalies (Chinese)Window readings and per-call ledger

Volcengine Coding Plan Pro · kimi-k3

Conditional estimate · low confidence · 2026-10-28

Native window: 5h · 30 calls · Used 0.57% → 10.01%

Volcengine Coding Plan Pro token usage
Source	Fresh input	Cache read	Output
Bench	1,237,470	0	21,228

Cache writes were zero; write weights are uncalibrated. Output includes reasoning without double-counting.

13,319,555–13,347,805 tokens / 5 hours

Used allowance: 0.57% → 10.01%

99,110,078–100,695,840 tokens / week

Used allowance: 12.07% → 13.33%

196,671,562–203,015,807 tokens / month (rough)

Used allowance: 33.71% → 34.34%

Each window measured directly. Ranges cover meter rounding only, not statistical confidence intervals.

30 calls in 3 batches, one model, about 98.31% fresh input and no cache reads/writes. Each window measured directly; monthly capacity is rough. Bounds cover meter rounding only and assume linear consumption without other account traffic; not a universal cross-model token cap.

Measurement process and calculations (Chinese)Meter readings and per-call usage

Other measured models

Same plan, measured separately per model. Conditional token ranges at each observed workload mix; monthly capacity is rough.

Model / calls	5 hours	Week	Month (rough)
deepseek-v4.1-flash24 calls · 96.75% fresh input	66.03M–66.92M	474.79M–524.77M	906.42M–1.11B
deepseek-v4-pro12 calls · 95.83% fresh input	18.13M–18.26M	132.61M–139.98M	251.96M–279.96M
glm-5.312 calls · 97.45% fresh input	18.12M–18.25M	136.13M–143.91M	265.1M–296.29M
glm-5.3-flash20 calls · 99.47% fresh input	132.78M–133.86M	993.83M–1.06B	1.93B–2.19B

Bounds cover meter rounding only. Output proportions differ, so these are not isolated model weights. Models share the allowance pool; capacities cannot be added.

GLM-5.3-Flash shows only the long-input profile (~163K input tokens per call). The first four shorter pilot calls are documented separately and excluded.

Multi-model methodology and mix (Chinese)Window readings and usage ledger

Cursor Ultra (legacy) · composer-2.5

Pending reconciliation · capacity withheld · 2026-10-28

Native window: monthly · 16 calls · Used 34.7423% → 34.757%

Cursor Ultra (legacy) token usage
Source	Fresh input	Cache read	Output
Bench	1,064,029	64,322	33,906

Cache writes were zero; write weights are uncalibrated. Output includes reasoning without double-counting.

Only the first 12 calls (801,760 tokens) belong to this meter interval; batch 4 has no final reading. Paused after browser disconnection; the 16-call sample lacks CSV reconciliation. Monthly window inferred from legacy two-pool rules; reset time not captured. Capacity withheld.

Full report (Chinese)Structured data JSON Sanitized usage ledger JSON