📊 Claude Token Usage Report2026-04-08

[github-actions]

Overview

Period: 2026-04-07T12:38Z to 2026-04-08T12:38Z (since previous report §1747)
Runs analyzed: 7 of 7 Secret Digger (Claude) runs — 6 had token data, 1 had no api-proxy-logs
Total tokens: 520K across 6 instrumented runs
Estimated total cost: $1.18

ℹ️ No Smoke Claude or Security Guard runs occurred in this period. Only Secret Digger (Claude) ran (7× scheduled, every ~57 min). Run §24121870764 had no api-proxy-logs directory in its firewall-audit-logs artifact — token data unavailable for that run.

---

Workflow Summary

Workflow	Runs	Total Tokens	Est. Cost	Cache Rate	I/O Ratio	Top Model
Secret Digger (Claude)	7 (6 w/ data)	520K	$1.18	100%*	30:1	claude-sonnet-4-6

*Cache hit rate of 100% on request 2 of each run; request 1 always writes fresh (0% read). Cross-run cache reuse is not possible at ~57-min intervals (Anthropic cache TTL is ~5 min).

Avg cost per run: $0.196 (vs $0.190 previous period, +3%)

---

🔍 Optimization Opportunities

1. Cache write dominates cost (~85%) — structurally unchanged

   • Each run writes 41,967-token system prompt to cache at $3.75/M = **$0.157/run** (~80% of total run cost)
   • Runs are ~57 minutes apart, far beyond Anthropic's 5-minute cache TTL — zero cross-run cache reuse is possible
   • Recommendation: No practical optimization given scheduling interval. Reducing frequency from ~1 hr to ~2 hr would save ~50% on cost.

2. Output tokens increased +29% vs previous period

   • Previous period avg: 1,094 output tokens/run; current avg: 1,410 output tokens/run
   • Despite this increase, total cost rose only marginally (+3%) because output cost is small relative to cache_write cost
   • No optimization needed, but worth watching if trend continues — output tokens represent actual reasoning work done.

3. One run (24121870764) had no api-proxy-logs

   • Run concluded failure at 06:44 UTC — the only run in this period without token tracking data
   • Most likely the api-proxy container failed to start or the artifact upload was incomplete
   • Recommendation: Investigate whether this is a one-off or recurring issue with api-proxy instrumentation.

4. 100% failure rate — all 7 runs concluded failure

   • Every Secret Digger (Claude) run detected prompt injection and filed a noop (same as all previous periods)
   • Recommendation: Verify whether failure conclusion is correct semantics for successful injection detection — it creates misleading CI signal.

---

Per-Run Details — Secret Digger (Claude)

All runs on branch main, head SHA 91bd673b, model claude-sonnet-4-6. Each run makes exactly 2 requests.

Run	Time (UTC)	Req 1 Output	Req 2 Output	Cache Write	Avg Latency	Cost
§24132929701	11:27	1,497	22	43,478	17,322ms	$0.1984
§24130802058	10:31	1,506	189	43,490	18,771ms	$0.2011
§24128477727	09:33	1,119	2	43,105	15,519ms	$0.1911
§24126082835	08:36	975	2	42,961	12,506ms	$0.1884
§24123839417	07:40	1,721	51	43,704	20,317ms	$0.2031
§24121870764	06:44	—	—	—	—	N/A (no data)
§24119865090	05:41	1,322	54	43,306	15,055ms	$0.1956

Token breakdown (6 runs with data):

• Input: 24 tokens total (~4 per run)
• Output: 8,460 tokens total (~1,410 per run)
• Cache read: 251,802 tokens total (~41,967 per run)
• Cache write: 260,044 tokens total (~43,341 per run)
• Full-price tokens (input + output + cache_write): 268,528 (~45K/run)

Cost breakdown per run (approx):

• Cache write (~43K tokens × $3.75/M): ~$0.161 (~82%)
• Output (~1,410 tokens × $15/M): ~$0.021 (~11%)
• Cache read (~42K tokens × $0.30/M): ~$0.013 (~7%)
• Input (~4 tokens × $3/M): ~$0.000 (<1%)

Historical Trend (last 6 reports)

Date	Runs	Tokens	Cost	Avg $/run	Avg Output/run	I/O Ratio
2026-04-08	6/7	520K	$1.18	$0.196	1,410	30:1
2026-04-07 §1747	7/7	603K	$1.33	$0.190	1,094	38:1
2026-04-06 §1719	7/7	~603K	~$1.33	~$0.190	—	—
2026-04-05 §1679	—	—	—	—	—	—
2026-04-04 §1670	—	—	—	—	—	—
2026-04-03 §1646	—	—	—	—	—	—

Trend assessment: Cost per run is stable (~$0.19–$0.20). Output tokens increased +29% this period (1,094 → 1,410) but had negligible cost impact. The system prompt cache (~42K tokens) is consistent across all runs, suggesting no changes to the Secret Digger prompt.

Previous Report

§1747 — 2026-04-07

References:

• §24135664930 — this analysis run
• §24132929701 — most recent Secret Digger (Claude)
• §1747 — previous Claude token usage report

Generated by Daily Claude Token Usage Analyzer · ● 1.7M · ◷