📊 Claude Token Usage Report2026-04-13

[github-actions]

Overview

Period: 2026-04-05 to 2026-04-13
Runs analyzed: 23 (all 23 had token data)
Total tokens: 4,305K (effective: 1,434K after Anthropic cache pricing)
Estimated total cost: $6.81
Total Actions minutes: 110 min
Total turns: 116

Three Claude-engine workflows ran: Secret Digger (Claude) (5 runs, highest per-run cost), Smoke Claude (10 runs, highest volume), and Security Guard (8 runs). All runs used claude-sonnet-4-6. Prompt caching is active across all workflows — but Secret Digger is the only workflow where cache writes exceed reads, indicating cache is not being reused across runs.

Workflow Summary

Workflow	Runs	Total Tokens	Avg Tokens	Total Cost	Avg Cost	Avg Turns	Errors
Secret Digger (Claude)	5	629K	126K	$2.55	$0.51	3.0	0
Smoke Claude	10	2,156K	216K	$2.21	$0.22	6.5	2
Security Guard	8	1,519K	190K	$2.06	$0.26	4.5	1
Total	23	4,305K	187K	$6.81	$0.30	5.0	3

🔍 Optimization Opportunities

1. Secret Digger (Claude) — $0.51/run, cache writes ≈ reads (0.97× ratio ⚠️)

   • Cache is being written each run but barely reused: 194,981 write tokens vs 189,635 read tokens across 5 runs
   • The high write/read parity suggests each run starts fresh without benefiting from prior cache entries
   • Consider whether system prompt caching is properly structured to reuse the stable prefix across runs — cache entries persist ~5 min on Anthropic; if runs are spaced further apart, cache won't be hit
   • Cache writes cost 12.5× more than reads; maximizing the read/write ratio is the biggest cost lever for this workflow

2. Smoke Claude — model_downgrade_available flagged on 9 of 10 runs

   • 8 of 10 runs were classified as "Triage" domain; the assessment recommends a smaller model (e.g. claude-haiku-4-5) could handle the task at lower cost
   • poor_agentic_control flagged on 5 runs — broad or weakly bounded tool access
   • resource_heavy_for_domain flagged on 5 runs
   • ~50–57% of turns are data-gathering that could move to deterministic pre-agent steps
   • Cache is healthy (4.13× read/write ratio) but avg 216K tokens/run is high for a triage workflow
   • 2 failure runs this period

3. Security Guard — model_downgrade_available flagged on 1 run; ~60% turns reducible

   • 1 of 8 runs assessed as "Issue Response" domain — likely doesn't require frontier model
   • ~50–60% of turns across affected runs are pure data-gathering and could be pre-computed
   • Cache ratio is solid (3.13×) — caching is working well here
   • 1 failure run this period

Per-Workflow Details

Secret Digger (Claude)

• Runs: 5 — 24012669686 ✅ · 24011656105 ✅ · 24010635315 ✅ · 24009579929 ✅ · 24008628142 ✅
• All runs: 2026-04-05 (5 hourly runs)
• Total tokens: 629K (avg 126K/run) | Effective: 239K
• Estimated cost: $2.55 (avg $0.51/run)
• Turns: 15 total (avg 3.0/run)
• Actions minutes: 29 min total
• Model: claude-sonnet-4-6
• Cache analysis:
  • Cache reads: 189,635 tokens
  • Cache writes: 194,981 tokens
  • Ratio: 0.97× ⚠️ — writes barely exceed reads; cache is not being reused
  • Raw input: 635 tokens | Output: 6,179 tokens
• Error rate: 0/5 runs

Smoke Claude

• Runs: 10 — 24345637612 ❌ · 24321359131 ❌ · 24319647106 ✅ · 24319183558 ✅ · 24318847644 ✅ · 24318792261 ✅ · 24318513152 ✅ · 24316781100 ✅ · 24315641999 ✅ · 24315473861 ✅
• Total tokens: 2,156K (avg 216K/run) | Effective: 623K
• Estimated cost: $2.21 (avg $0.22/run)
• Turns: 65 total (avg 6.5/run)
• Actions minutes: 44 min total
• Model: claude-sonnet-4-6
• Cache analysis:
  • Cache reads: 1,729,202 tokens
  • Cache writes: 418,778 tokens
  • Ratio: 4.13× ✅ — cache is working well
  • Raw input: 67 tokens | Output: 7,774 tokens
• Agentic flags: partially_reducible (10/10), resource_heavy_for_domain (5/10), poor_agentic_control (5/10), model_downgrade_available (9/10)
• Error rate: 2/10 runs

Security Guard

• Runs: 8 — 24319647097 ✅ · 24319183550 ✅ · 24318847668 ✅ · 24318792239 ✅ · 24318513148 ✅ · 24316781082 ✅ · 24315642001 ❌ · 24315473875 ✅
• Total tokens: 1,519K (avg 190K/run) | Effective: 572K
• Estimated cost: $2.06 (avg $0.26/run)
• Turns: 36 total (avg 4.5/run)
• Actions minutes: 37 min total
• Model: claude-sonnet-4-6
• Cache analysis:
  • Cache reads: 1,132,847 tokens
  • Cache writes: 362,336 tokens
  • Ratio: 3.13× ✅ — cache is healthy
  • Raw input: 48 tokens | Output: 24,152 tokens (highest output of all workflows)
• Agentic flags: partially_reducible (2/8), model_downgrade_available (1/8)
• Error rate: 1/8 runs

Workflows Without Token Data

No workflows had missing token data in this report period — all 23 runs had complete token instrumentation via the api-proxy.

Historical Trend

Compared to the previous report (#1951 — 2026-04-12):

• Previous: 50 runs · $16.86 total · 226 turns · 250 action min
• Current: 23 runs · $6.81 total · 116 turns · 110 action min

The lower totals reflect a shorter observation window and fewer Secret Digger runs (5 vs 21 in the prior period). Average per-run costs are comparable: Smoke Claude ~$0.22/run (stable), Security Guard ~$0.26/run (stable). The Secret Digger cache problem (cache writes ≈ reads) is newly visible this period and warrants investigation.

Previous Report

#1951 — 2026-04-12

References:

• §24364461591 — this report's workflow run

Generated by Daily Claude Token Usage Analyzer · ● 356.4K · ◷

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #1987

View newer issue

---

This action was performed automatically by the Daily Claude Token Usage Analyzer workflow.