🇰🇷 한국어 버전
Date: April 8, 2026 — 110 main sessions + 279 subagent sessions (JSONL, April 1-6), v2.1.91, Max 20x ($200/mo)
Relationship to other documents: 02_RATELIMIT-HEADERS.md analyzes server-side rate limit headers from the proxy. This document analyzes client-side JSONL session logs (
~/.claude/projects/**/*.jsonl). Section 6 cross-correlates JSONL with proxy data. Full proxy dataset and JSONL bulk scan: 13_PROXY-DATA.md.
JSONL session logs are the client's record of every API interaction. Unlike the proxy (which sees HTTP response headers), JSONL captures per-turn token breakdowns including PRELIM/FINAL entries, synthetic rate limit events, and model identification.
Key findings:
- Synthetic rate limiting (B3): 24 confirmed
<synthetic>entries across 6 days (JSONL). At corpus scale: 183/532 files (34.4%) contain synthetic markers, 3,129 rate limit text occurrences across 257 files (48.3%). See 13_PROXY-DATA.md §12 - PRELIM inflation (B8): 5,551 PRELIM vs 6,825 FINAL entries (0.81x ratio overall, up to 1.14x on heavy days). Bulk scan (532 files): average inflation 2.37x, worst case 4.42x. See 13_PROXY-DATA.md §12.3
- Subagent overhead: 279 subagent sessions consumed 17.3% of total cache_read but 40.4% of total output. Cold start median: 13,358 cache_create tokens. Proxy confirms: 3,781 haiku requests at 58.1% cache vs 11,959 opus at 98.8%
- Session growth: cache_read grows 24x over a 990-turn session (25K → 595K per turn, +575 linear). Proxy (53 sessions): median 1,845 tok/min (P25: 801, P75: 3,581). See 13_PROXY-DATA.md §3
- Session cost range: 753x variation (447K → 336M visible tokens) between cheapest and most expensive sessions
- Time-normalized cost: median 153K cache_read/minute, mean 227K/minute
- Cross-correlation with proxy: JSONL records 1.93x the cache_read tokens that the proxy sees — directly confirming PRELIM double-counting inflates local accounting by ~2x
- Budget enforcement (proxy): 167,818 events across 218 sessions (April 1–13) — 100% truncation rate, 100% ≤50 chars. See 13_PROXY-DATA.md §7
- Microcompact (proxy): 5,500 events, 18,858 items cleared (April 1–13). Scales with message count: 1.6 cleared at <10 msgs → 6.6 at 200+. See 13_PROXY-DATA.md §8
Main sessions only (excludes subagents):
| Day | Sessions | Entries | Output | Cache Read | Cache Create | Input | Total Visible |
|---|---|---|---|---|---|---|---|
| 04-01 | 25 | 1,930 | 567,657 | 173,259,918 | 21,418,689 | 72,855 | 195,319,119 |
| 04-02 | 29 | 3,014 | 826,510 | 447,692,321 | 8,155,349 | 229,115 | 456,903,295 |
| 04-03 | 15 | 1,203 | 260,041 | 84,453,778 | 2,712,374 | 26,520 | 87,452,713 |
| 04-04 | 21 | 2,972 | 782,955 | 531,652,242 | 7,692,177 | 187,139 | 540,314,513 |
| 04-05 | 12 | 1,426 | 316,962 | 132,464,357 | 6,455,937 | 39,222 | 139,276,478 |
| 04-06 | 9 | 1,858 | 501,087 | 353,322,699 | 4,331,168 | 33,201 | 358,188,155 |
Per-entry averages:
| Day | Cache Read % | Create % | Output/Entry | Cache/Entry | Total/Entry |
|---|---|---|---|---|---|
| 04-01 | 88.7% | 11.0% | 294 | 89,772 | 101,202 |
| 04-02 | 98.0% | 1.8% | 274 | 148,538 | 151,594 |
| 04-03 | 96.6% | 3.1% | 216 | 70,203 | 72,696 |
| 04-04 | 98.4% | 1.4% | 263 | 178,887 | 181,802 |
| 04-05 | 95.1% | 4.6% | 222 | 92,892 | 97,669 |
| 04-06 | 98.6% | 1.2% | 270 | 190,163 | 192,782 |
Observations:
- Cache Read % ranges 88.7–98.4%. April 1 is lower (88.7%) due to many session cold starts (25 sessions in one day → more cache_create).
- Output per entry is only 216–294 tokens — this is visible output only. With thinking enabled, actual server-side output per turn is much higher.
- Cache per entry ranges 70K–179K. Heavier days (04-04) have longer sessions → more accumulated context per turn.
| Day | Synthetic (B3) | PRELIM | FINAL | P/F Ratio |
|---|---|---|---|---|
| 04-01 | 4 | 911 | 1,015 | 0.90 |
| 04-02 | 4 | 1,399 | 1,611 | 0.87 |
| 04-03 | 3 | 556 | 643 | 0.86 |
| 04-04 | 4 | 1,229 | 1,737 | 0.71 |
| 04-05 | 9 | 754 | 663 | 1.14 |
| 04-06 | 0 | 702 | 1,156 | 0.61 |
| Total | 24 | 5,551 | 6,825 | 0.81 |
- Synthetic (B3): 24 events. April 5 had 9 — the highest concentration. Each represents a client-generated fake "Rate limit reached" with no API call.
- PRELIM/FINAL ratio varies 0.61–1.14 depending on thinking usage. April 5 was 1.14x (more PRELIM than FINAL — heavy thinking). April 6 was 0.61x (lighter thinking).
Do PRELIM entries carry the same token counts as FINAL entries?
| Day | PRELIM cache_read | FINAL cache_read | P/F Cache Ratio | PRELIM output | FINAL output | P/F Output Ratio |
|---|---|---|---|---|---|---|
| 04-01 | 73,581,649 | 99,678,269 | 0.74 | 22,329 | 545,328 | 0.04 |
| 04-02 | 187,693,131 | 259,999,190 | 0.72 | 30,416 | 796,094 | 0.04 |
| 04-03 | 34,939,821 | 49,362,008 | 0.71 | 14,217 | 237,824 | 0.06 |
| 04-04 | 191,259,356 | 340,298,839 | 0.56 | 24,726 | 742,229 | 0.03 |
| 04-05 | 65,313,689 | 67,150,668 | 0.97 | 22,956 | 294,006 | 0.08 |
| 04-06 | 135,483,323 | 217,839,376 | 0.62 | 13,028 | 488,059 | 0.03 |
PRELIM cache_read is 0.56–0.97x of FINAL cache_read — substantial but not 1:1. The ratio varies by day. PRELIM output is only 0.03–0.08x of FINAL (PRELIM entries have almost no visible output — they're intermediate "thinking in progress" snapshots).
This means: PRELIM inflation (B8) adds roughly 56–97% additional cache_read on top of the real FINAL values. If the server counts PRELIM entries at full weight, users are paying 1.56–1.97x for cache_read.
Not all sessions are equal. The 753x range between cheapest and most expensive sessions shows how dramatically costs vary:
| Day | Session | Total Visible | Entries | Cache Read | Output |
|---|---|---|---|---|---|
| 04-04 | 03480ffe | 336,360,170 | 990 | 333,169,783 | 266,151 |
| 04-02 | 63e79b5e | 163,952,464 | 609 | 161,907,520 | 155,012 |
| 04-06 | b926be34 | 126,541,892 | 460 | 124,243,773 | 180,303 |
| 04-06 | ae47b46b | 102,192,853 | 694 | 101,261,459 | 83,175 |
| 04-02 | 158a1b9d | 69,083,572 | 354 | 68,264,817 | 98,020 |
| 04-01 | 888b0e4b | 60,747,270 | 258 | 59,511,929 | 123,305 |
| 04-02 | 89bdbe06 | 48,365,855 | 327 | 47,806,054 | 100,093 |
| 04-02 | 23d083a1 | 42,169,722 | 265 | 41,368,814 | 94,055 |
| 04-02 | 48557289 | 40,430,993 | 309 | 39,400,900 | 69,508 |
| 04-04 | 0c024075 | 39,933,209 | 284 | 39,283,156 | 69,145 |
| Day | Session | Total Visible | Entries | Cache Read | Output |
|---|---|---|---|---|---|
| 04-05 | 41549298 | 446,988 | 12 | 419,350 | 2,685 |
| 04-06 | fb0c85e5 | 463,571 | 12 | 430,191 | 2,093 |
| 04-02 | 79388f46 | 474,370 | 13 | 429,085 | 5,073 |
| 04-02 | 22b91f19 | 501,069 | 13 | 401,350 | 4,903 |
| 04-03 | 30f6ba7a | 587,411 | 17 | 538,729 | 4,652 |
| Metric | Value |
|---|---|
| Sessions analyzed | 78 (with >10 entries and >1 min duration) |
| Cost range | 446,988 — 336,360,170 (753x) |
| Median | 8,581,558 |
| Mean | 21,620,755 |
The most expensive session (03480ffe, 990 turns, 336M visible tokens) consumed as much as 38 median sessions combined. This is the natural consequence of cache_read growing linearly with conversation length — long sessions are disproportionately expensive.
Raw token totals can be misleading when comparing sessions or days with different work hours. Time-normalized metrics give a fairer picture:
| Metric | Value |
|---|---|
| Median duration | 49 min (0.8h) |
| Mean duration | 158 min (2.6h) |
| Min | 1 min |
| Max | 1,372 min (22.9h) |
| Metric | Value |
|---|---|
| Median | 153,018 tokens/min |
| Mean | 226,675 tokens/min |
| Min | 8,397 tokens/min |
| Max | 864,186 tokens/min |
| Metric | Value |
|---|---|
| Median | 574 tokens/min |
| Mean | 744 tokens/min |
Why this matters: When comparing budget reduction across time periods, total token consumption includes both "how long you worked" and "how expensive each minute was." A day with 16h work and 2.7B cache_read (168M/h) vs a day with 2h work and 123M cache_read (62M/h) shows 22x total reduction but only 2.7x per-hour reduction. The per-hour metric isolates the actual cost change from the work-time difference.
This is relevant to community analyses (e.g., @fgrosswig's comparison of March 30 vs April 6) where total-to-total comparisons can overstate budget reduction by conflating work hours with cost-per-hour changes.
pie title Token Share: Main vs Subagent
"Main Cache Read (83%)" : 1607617725
"Subagent Cache Read (17%)" : 335964567
pie title Cache Create Share
"Main (38%)" : 50302333
"Subagent (62%)" : 82034105
| Metric | Main Sessions | Subagents | Subagent Share |
|---|---|---|---|
| Files | 110 | 279 | — |
| Entries | 11,983 | 10,178 | — |
| Output | 3,100,588 | 2,100,541 | 40.4% |
| Cache Read | 1,607,617,725 | 335,964,567 | 17.3% |
| Cache Create | 50,302,333 | 82,034,105 | 62.0% |
| Input | 564,967 | 3,477,557 | — |
| Cache Read % | 96.9% | 79.7%* | — |
| Avg cache_read/entry | 134,158 | 33,009 | 4.1x |
*79.7% is a per-entry weighted average (total subagent cache_read / total subagent visible tokens), not a simple mean of per-session ratios. Sessions with more entries contribute proportionally more to this figure.
| Metric | Value |
|---|---|
| Samples | 279 subagent sessions |
| Median cache_create (1st turn) | 13,358 tokens |
| Mean | 12,105 |
| P25 — P75 | 7,205 — 16,703 |
| Min — Max | 491 — 25,075 |
Each subagent spawn costs roughly 13K tokens in cache_create on the first turn. With 279 subagent sessions over 6 days, that's ~3.7M tokens spent just on cold starts.
| Metric | Value |
|---|---|
| Samples | 9,403 entries (turn 4+) |
| Median cache_read % | 92.3% |
| Mean | 81.7% |
| P25 — P75 | 78.1% — 97.6% |
| Min — Max | 0.0% — 100.0% |
After 3+ turns, subagent cache_read settles to a median 92.3% — lower than main sessions (96.9%) but functional. The P25 of 78.1% indicates some subagents never fully warm, likely due to short lifespans (many subagents run <5 turns).
What this means:
- Subagents produce 40% of output but only consume 17% of cache_read — each turn is lighter (33K vs 134K) because subagents have shorter context.
- Subagents account for 62% of cache_create — they frequently start cold. This is the expensive operation at ~13K tokens per spawn.
- Users with heavy subagent workflows (e.g., many Agent tool calls) pay disproportionately in cache_create but save on cache_read per turn.
xychart-beta
title "Cache Read Growth (session 03480ffe, 990 turns)"
x-axis ["1","10","50","100","200","281","301","500","700","990"]
y-axis "Cache Read (K tokens)" 0 --> 600
line [25,34,64,145,204,8,257,347,439,595]
Analysis of the longest session (03480ffe, 990 entries, 336M visible tokens):
| Turn | Cache Read | Growth from Turn 1 | Cache Read % |
|---|---|---|---|
| 1 | 24,820 | — | 76.2% |
| 10 | 34,131 | 1.4x | 94.3% |
| 50 | 63,942 | 2.6x | 99.0% |
| 100 | 145,104 | 5.8x | 99.3% |
| 200 | 204,247 | 8.2x | 99.1% |
| 281 | 8,237 | 0.3x (anomaly) | 3.4% |
| 301 | 257,309 | 10.4x | 97.4% |
| 500 | 347,470 | 14.0x | 99.8% |
| 700 | 439,024 | 17.7x | 99.8% |
| 990 | 595,141 | 24.0x | 100.0% |
Growth is linear — roughly +575 cache_read tokens per turn. This is conversation history accumulating. At turn 990, each API call sends ~595K tokens of cached context.
Anomaly at turn 281: cache_read dropped from ~222K to 8,237 with cache_create spiking to 237,490. This is a mid-session cache rebuild — possibly triggered by microcompact (B4) or a server-side cache eviction. Recovery to pre-anomaly levels took ~20 turns.
Implication for utilization: If each 1% of 5h utilization costs ~1.5M-2.1M visible tokens (from 02_RATELIMIT-HEADERS.md), then:
- Early session (turn 10, ~34K cache/turn): ~44-62 turns per 1%
- Mid session (turn 500, ~347K cache/turn): ~4-6 turns per 1%
- Late session (turn 990, ~595K cache/turn): ~2.5-3.5 turns per 1%
A 100-turn stretch in the late half of a long session costs roughly 17-40% of the 5h window in cache_read alone. This is why long sessions drain faster — the per-turn cost accelerates even though the user's work pace stays constant.
Both data sources cover the same API calls (April 4-6) but measure different things:
- JSONL: logs what the client recorded, including PRELIM entries
- Proxy: logs what the server returned in a single HTTP response
By comparing hourly totals from both sources (33 overlapping hours):
| Metric | JSONL Total | Proxy Total | Ratio |
|---|---|---|---|
| Output tokens | 1,372,025 | 1,968,953 | 0.70x |
| Cache Read tokens | 663,021,317 | 344,380,772 | 1.93x |
This is the direct, measured confirmation of Bug 8 (PRELIM inflation). The JSONL records PRELIM entries with comparable (0.56-0.97x) cache_read_input_tokens to the FINAL entry, effectively inflating the count. The ~2x ratio matches the PRELIM/FINAL token analysis from Section 1 (PRELIM cache is 0.56-0.97x of FINAL, so total JSONL = FINAL + PRELIM ≈ 1.56-1.97x of FINAL-only, consistent with 1.93x).
JSONL records fewer output tokens than the proxy. This is primarily a timing/matching artifact — subagent JSONL files have different modification times than when the proxy logged the corresponding request, causing some entries to fall outside the matched hourly window.
- If you sum
cache_read_input_tokensfrom your JSONL files, your total will be ~2x higher than what was actually sent to the API — because PRELIM entries duplicate the count. - If Anthropic's server-side rate limiter uses the same accounting that produces PRELIM entries, users could be charged double for cache_read. We cannot confirm this from the client side, but the 1.93x ratio is consistent with the hypothesis.
When comparing token totals across different time periods, always normalize by work hours or active minutes. Total-to-total comparisons conflate "how long you worked" with "how expensive each minute was." For example, 16h × 168M/h = 2.7B total vs 2h × 62M/h = 123M total gives a 22x total difference, but the per-hour cost difference is only 2.7x. See Section 3 for time-normalized metrics from our data.
- JSONL timestamps: Not all entries have precise timestamps. Cross-correlation uses hourly buckets, which introduces matching error. Individual hours can show extreme ratios (35x+) due to timing misalignment.
- Subagent attribution: Subagent JSONL files are stored by parent session. Modification times may not reflect actual request times.
- No thinking token visibility: Like the proxy, JSONL
output_tokensdoes not include extended thinking tokens. The actual server-side output per turn is higher than recorded. - 6-day window: April 1-6 only. Longer observation would reveal weekly patterns.
- Single plan tier: Max 20x ($200/mo). Different tiers may have different per-entry overhead characteristics.
- PRELIM detection heuristic: PRELIM entries are identified by empty/null
stop_reason. Other entry types with empty stop_reason could be miscounted.
| Question | JSONL Answer | Proxy Answer | Combined |
|---|---|---|---|
| Cache_read per turn? | 134K main, 33K subagent | — | ✓ Measured |
| Cache growth per turn? | +575 tokens/turn (linear) | — | ✓ Measured |
| Cache_read per minute? | median 153K/min | — | ✓ Measured |
| What does 1% utilization cost? | — | 1.5M-2.1M visible | ✓ Measured |
| Are PRELIM entries inflating counts? | Yes (0.81x P/F ratio) | Yes (JSONL is 1.93x of proxy) | ✓ Confirmed from both sides |
| PRELIM cache vs FINAL cache? | 0.56-0.97x (not 1:1) | — | ✓ Measured |
| Are synthetic rate limits real? | 24 events in 6 days | — | ✓ Measured |
| Subagent cold start cost? | median 13,358 cache_create | — | ✓ Measured |
| Subagent warm cache %? | median 92.3% (turn 4+) | — | ✓ Measured |
| Session cost range? | 753x (447K-336M visible) | — | ✓ Measured |
| Are thinking tokens counted? | Can't see them | Can't see them | Still unknown |
| Server token-type weights? | — | Can't decompose | Still unknown |
The two remaining unknowns — thinking token accounting and server-side token weighting — require either Anthropic disclosure or the thinking-disabled isolation test planned for this week.
Hours with >50 entries shown. Full dataset covers April 1-6, 2026 (UTC).
| Hour (UTC) | Entries | Output | Cache Read | Cache Create | Synthetic | PRELIM | FINAL |
|---|---|---|---|---|---|---|---|
| 04-01T00 | 389 | 130,065 | 35,234,018 | 716,913 | 0 | 191 | 198 |
| 04-01T01 | 305 | 94,931 | 43,027,056 | 913,611 | 0 | 154 | 151 |
| 04-01T02 | 204 | 72,350 | 34,744,866 | 269,215 | 0 | 76 | 128 |
| 04-01T03 | 79 | 16,947 | 13,632,765 | 1,695,887 | 0 | 31 | 48 |
| 04-01T04 | 168 | 64,094 | 11,834,661 | 1,162,842 | 0 | 75 | 93 |
| 04-01T05 | 112 | 23,925 | 2,991,981 | 8,394,712 | 0 | 51 | 61 |
| 04-01T06 | 53 | 10,304 | 591,448 | 3,695,356 | 1 | 19 | 33 |
| 04-01T08 | 78 | 10,570 | 1,027,516 | 1,991,228 | 0 | 34 | 44 |
| 04-01T09 | 130 | 36,142 | 6,856,721 | 643,068 | 1 | 61 | 68 |
| 04-01T10 | 189 | 41,809 | 12,846,688 | 404,192 | 0 | 101 | 88 |
| 04-01T11 | 69 | 16,724 | 10,000,558 | 84,459 | 0 | 29 | 40 |
| 04-01T12 | 83 | 25,008 | 7,627,403 | 158,323 | 0 | 42 | 41 |
| 04-01T22 | 66 | 15,790 | 10,961,407 | 42,139 | 0 | 14 | 52 |
| 04-01T23 | 133 | 37,583 | 6,282,145 | 471,344 | 0 | 86 | 47 |
| 04-02T00 | 506 | 155,854 | 57,797,245 | 1,255,725 | 0 | 261 | 245 |
| 04-02T01 | 417 | 105,009 | 57,478,025 | 728,877 | 0 | 218 | 199 |
| 04-02T02 | 531 | 170,540 | 112,841,075 | 2,269,831 | 3 | 238 | 290 |
| 04-02T03 | 263 | 54,035 | 67,586,455 | 251,630 | 0 | 102 | 161 |
| 04-02T04 | 386 | 108,469 | 49,008,927 | 1,184,661 | 1 | 182 | 203 |
| 04-02T05 | 247 | 60,096 | 38,093,933 | 206,106 | 0 | 90 | 157 |
| 04-02T06 | 52 | 12,946 | 8,804,049 | 52,356 | 0 | 15 | 37 |
| 04-02T15 | 301 | 59,421 | 23,193,271 | 563,409 | 1 | 135 | 165 |
| 04-02T16 | 58 | 8,255 | 4,685,862 | 138,018 | 0 | 28 | 30 |
| 04-02T17 | 206 | 48,828 | 16,995,785 | 204,872 | 0 | 68 | 138 |
| 04-02T21 | 97 | 13,429 | 4,071,685 | 233,632 | 0 | 55 | 42 |
| 04-02T22 | 195 | 39,381 | 11,892,785 | 558,199 | 1 | 106 | 88 |
| 04-02T23 | 102 | 20,803 | 5,084,026 | 211,964 | 0 | 55 | 47 |
| 04-03T00 | 216 | 66,913 | 24,476,045 | 541,620 | 0 | 100 | 116 |
| 04-03T01 | 330 | 84,131 | 78,646,938 | 2,461,625 | 5 | 144 | 181 |
| 04-03T02 | 248 | 60,545 | 98,032,051 | 199,657 | 0 | 86 | 162 |
| 04-03T03 | 90 | 22,274 | 43,862,998 | 64,566 | 0 | 21 | 69 |
| 04-03T04 | 122 | 31,699 | 65,962,917 | 84,870 | 0 | 36 | 86 |
| 04-03T06 | 166 | 40,002 | 24,972,321 | 224,897 | 0 | 60 | 106 |
| 04-03T14 | 76 | 20,497 | 5,369,588 | 120,679 | 0 | 18 | 58 |
| 04-03T15 | 124 | 48,573 | 10,578,844 | 198,689 | 0 | 69 | 54 |
| 04-04T00 | 60 | 13,593 | 4,008,263 | 619,547 | 0 | 32 | 28 |
| 04-04T01 | 218 | 45,017 | 15,814,301 | 574,946 | 0 | 86 | 132 |
| 04-04T02 | 252 | 58,591 | 34,358,168 | 440,197 | 0 | 135 | 117 |
| 04-04T03 | 314 | 65,399 | 27,765,417 | 528,688 | 0 | 158 | 156 |
| 04-04T04 | 79 | 21,098 | 4,584,805 | 158,716 | 0 | 36 | 43 |
| 04-04T05 | 141 | 18,514 | 8,892,255 | 238,306 | 0 | 66 | 75 |
| 04-04T06 | 72 | 11,050 | 7,547,155 | 42,904 | 0 | 33 | 39 |
| 04-04T11 | 124 | 32,170 | 15,136,178 | 397,115 | 0 | 48 | 76 |
| 04-04T12 | 54 | 34,569 | 3,559,586 | 121,601 | 0 | 17 | 36 |
| 04-04T13 | 61 | 20,215 | 5,390,607 | 153,114 | 0 | 19 | 42 |
| 04-04T14 | 328 | 102,198 | 48,194,992 | 460,172 | 0 | 118 | 210 |
| 04-04T22 | 283 | 49,910 | 16,432,240 | 1,617,418 | 2 | 177 | 104 |
| 04-04T23 | 121 | 25,556 | 12,158,889 | 138,953 | 0 | 61 | 60 |
| 04-05T04 | 53 | 10,285 | 2,463,098 | 332,913 | 0 | 37 | 16 |
| 04-05T05 | 453 | 92,991 | 47,579,878 | 1,665,956 | 1 | 191 | 261 |
| 04-05T06 | 168 | 37,740 | 18,376,958 | 305,179 | 0 | 75 | 93 |
| 04-05T09 | 158 | 39,929 | 13,943,379 | 681,542 | 1 | 73 | 84 |
| 04-05T10 | 176 | 36,773 | 16,840,145 | 1,048,552 | 1 | 84 | 91 |
| 04-05T12 | 157 | 34,253 | 17,515,016 | 510,358 | 3 | 82 | 72 |
| 04-05T13 | 52 | 18,513 | 5,033,990 | 51,576 | 0 | 11 | 41 |
| 04-05T14 | 122 | 32,031 | 11,027,184 | 793,276 | 1 | 53 | 68 |
| 04-05T15 | 153 | 47,630 | 21,898,419 | 252,159 | 0 | 71 | 82 |
| 04-05T22 | 91 | 31,861 | 9,890,214 | 201,216 | 0 | 47 | 44 |
| 04-05T23 | 50 | 11,787 | 8,004,948 | 244,134 | 0 | 15 | 35 |
| 04-06T00 | 149 | 67,941 | 27,682,717 | 298,649 | 0 | 60 | 89 |
| 04-06T01 | 144 | 39,822 | 38,870,639 | 154,155 | 0 | 54 | 90 |
| 04-06T02 | 173 | 34,229 | 45,493,247 | 149,881 | 0 | 73 | 100 |
| 04-06T03 | 53 | 8,116 | 7,646,424 | 302,047 | 0 | 21 | 32 |
| 04-06T04 | 349 | 105,626 | 91,581,938 | 1,641,011 | 0 | 147 | 202 |
Rate limit headers captured via cc-relay proxy. Each row = one clock hour (KST).
| Hour (KST) | Reqs | Output | Cache Read | Input | 5h util | 7d util |
|---|---|---|---|---|---|---|
| 04-04 09 | 74 | 34,368 | 2,523,601 | 244,743 | 4% | 6% |
| 04-04 10 | 304 | 154,768 | 12,224,810 | 1,092,266 | 16% | 8% |
| 04-04 11 | 410 | 179,010 | 27,317,171 | 48,757 | 23% | 9% |
| 04-04 12 | 388 | 158,247 | 23,770,875 | 62,670 | 32% | 10% |
| 04-04 13 | 156 | 91,357 | 3,577,083 | 649,641 | 39% | 11% |
| 04-04 14 | 77 | 16,938 | 4,778,807 | 4,100 | 2% | 11% |
| 04-04 15 | 39 | 10,282 | 4,129,802 | 49 | 2% | 11% |
| 04-04 18 | 16 | 6,503 | 1,807,402 | 20 | 4% | 11% |
| 04-04 20 | 93 | 39,232 | 9,215,178 | 69,206 | 4% | 12% |
| 04-04 21 | 89 | 61,574 | 3,441,361 | 435,855 | 9% | 12% |
| 04-04 22 | 80 | 46,656 | 5,023,044 | 31,703 | 11% | 13% |
| 04-04 23 | 371 | 212,743 | 37,866,810 | 32,748 | 27% | 15% |
| 04-05 07 | 13 | 4,484 | 317,744 | 4,990 | 5% | 15% |
| 04-05 08 | 38 | 13,034 | 2,042,313 | 600 | 8% | 16% |
| 04-05 13 | 19 | 9,367 | 905,905 | 26 | 1% | 16% |
| 04-05 14 | 375 | 161,991 | 29,579,110 | 179,609 | 17% | 18% |
| 04-05 15 | 198 | 126,249 | 12,448,681 | 365,714 | 26% | 19% |
| 04-05 16 | 24 | 1,979 | 2,463,947 | 40 | 26% | 19% |
| 04-05 18 | 165 | 65,864 | 9,575,191 | 162,835 | 7% | 20% |
| 04-05 19 | 167 | 73,519 | 10,659,315 | 251,812 | 15% | 21% |
| 04-05 21 | 90 | 34,033 | 8,511,367 | 23,111 | 20% | 22% |
| 04-05 22 | 44 | 19,008 | 3,259,353 | 69 | 21% | 22% |
| 04-05 23 | 200 | 115,995 | 12,659,554 | 112,454 | 13% | 24% |
| 04-06 00 | 112 | 69,549 | 13,537,745 | 213 | 18% | 25% |
| 04-06 01 | 18 | 1,335 | 2,288,596 | 29 | 18% | 25% |
| 04-06 02 | 18 | 1,386 | 2,673,247 | 33 | 18% | 25% |
| 04-06 03 | 20 | 1,811 | 2,924,441 | 34 | 18% | 25% |
| 04-06 04 | 18 | 1,510 | 2,007,438 | 28 | 0% | 25% |
| 04-06 05 | 18 | 1,538 | 2,750,357 | 32 | 0% | 25% |
| 04-06 06 | 25 | 7,601 | 3,291,379 | 453 | 1% | 25% |
| 04-06 07 | 43 | 15,252 | 2,959,928 | 373 | 2% | 25% |
Environment: dataset ubuntu-1-stock — Max 20x ($200/mo), Opus 4.6 1M, v2.1.91, Linux (ubuntu-1), native ~/.claude (CC stock mode). JSONL corpus at the time of authoring: 1,735 files, 1.0 GB. Proxy cross-correlation subset: 8,794 requests, 3,702 with rate limit headers (April 4-6). Full proxy dataset (17,610 requests, April 1-8) and JSONL bulk scan (532 files): 13_PROXY-DATA.md. A parallel ubuntu-1-override environment (isolated override environment with a GrowthBook flag override active since April 10, same account) is tracked separately — see 14_DATA-SOURCES.md for the full label matrix and current totals.