📊 Claude Token Usage Report2026-04-11

[github-actions]

Overview

Period: 2026-04-05 to 2026-04-11 (7 days)
Runs analyzed: 50 (44 had token data; 6 cancelled with no data)
Total tokens: 9,238K across 3 workflows
Effective tokens: 2,989K (after Anthropic cache pricing adjustment)
Estimated total cost: $12.39
Total Actions minutes: 232 min
Total turns: 252

Three Claude-engine workflows ran: Security Guard (20 runs), Smoke Claude (23 runs), and Secret Digger (Claude) (7 runs). Overall costs declined significantly vs the previous report — Security Guard avg cost dropped 53% and Smoke Claude dropped 42%. Secret Digger (Claude) continues to exhibit a systemic cache inefficiency where cache writes exceed cache reads on every run.

---

Workflow Summary

Workflow	Runs	Total Tokens	Avg Tokens	Total Cost	Avg Cost	Avg Turns	Errors	Cache Reuse
Security Guard	20	3,731K	187K	$4.47	$0.22	4.8	8 (40%) ⚠️	82% ✅
Smoke Claude	23	4,623K	201K	$4.31	$0.19	5.8	6 (26%) ⚠️	83% ✅
Secret Digger (Claude)	7	884K	126K	$3.60	$0.51	3.0	0 (0%) ✅	49% ⚠️
Total	50	9,238K	—	$12.39	$0.25	5.0	14 (28%)	—

---

🔍 Optimization Opportunities

1. Secret Digger (Claude) — Cache write ≥ cache read on every run (49% reuse) ⚠️

   • All 7 runs this period show cache_write ≈ cache_read (~38K each), meaning the cache is created fresh each run and never hit from a prior run. Cross-run caching is not occurring.
   • With only 3 turns per run and a short context window, there is minimal intra-run reuse as well.
   • Recommendation: Move large static context (system prompt, repository facts, reference data) into a deterministic pre-agent step and pass via /tmp/gh-aw/agent/ files. This reduces expensive cache-write tokens each run. Consider consolidating runs or scheduling more frequently to improve cross-run cache hit rate.

2. Security Guard — High error rate (8/20 = 40%) ⚠️

   • 8 runs with errors is the highest error rate of any workflow. This drives unnecessary retry turns and inflates token costs.
   • Two outlier runs consumed disproportionate tokens: §24290111812 (480K tokens, 11 turns, $0.68) and §24271368232 (981K tokens, 24 turns, $0.57).
   • Recommendation: Investigate root causes of the 8 erroring runs. The 24-turn outlier suggests an agent retry loop — add a max_turns guardrail or improve error handling to prevent runaway turns.

3. Smoke Claude — Moderate error rate (6/23 = 26%) ⚠️

   • Approaching the 30% error threshold. Each erroring run wastes tokens on incomplete work.
   • Recommendation: Review error patterns to determine if these are infrastructure failures or agent logic errors. If systematic, consider adding retry logic or pre-flight checks.

4. Security Guard — High token variance

   • Avg is 187K/run but two runs exceeded 480K and 981K. High variance suggests the workflow is sensitive to input size or agent path choices.
   • Recommendation: Add input size guards or deterministic pre-processing steps to bound the agent's search space.

---

Per-Workflow Details

Security Guard

• Runs: 20 (§24290215023, §24290111812, §24289182705, §24289175288, §24287824916, and 15 more)
• Total tokens: 3,731K (avg 187K/run)
• Effective tokens: 1,461K
• Estimated cost: $4.47 (avg $0.22/run)
• Turns: 97 total (avg 4.8/run)
• Model: claude-sonnet-4-6 (Anthropic)
• Cache analysis: 3,693K read / 817K write → 82% reuse rate ✅
• Error rate: 8/20 runs (40%) ⚠️ — includes 6 cancelled runs with no token data
• Notable outliers:
  • §24290111812: 480K tokens, 11 turns, $0.68 (most expensive run in period)
  • §24271368232: 981K tokens, 24 turns, $0.57 (most turns in period — potential retry loop)

Smoke Claude

• Runs: 23 (§24290111810, §24290215008, §24289182699, §24289175294, §24287824928, and 18 more)
• Total tokens: 4,623K (avg 201K/run)
• Effective tokens: 1,186K
• Estimated cost: $4.31 (avg $0.19/run)
• Turns: 134 total (avg 5.8/run)
• Model: claude-sonnet-4-6 (Anthropic)
• Cache analysis: 3,789K read / 749K write → 83% reuse rate ✅
• Error rate: 6/23 runs (26%) ⚠️

Secret Digger (Claude)

• Runs: 7 (§24012669686, §24011656105, §24010635315, §24009579929, §24008628142, §24007580596, §24006466570)
• Total tokens: 884K (avg 126K/run)
• Effective tokens: 343K
• Estimated cost: $3.60 (avg $0.51/run)
• Turns: 21 total (avg 3.0/run)
• Model: claude-sonnet-4-6 (Anthropic)
• Cache analysis: 265K read / 274K write → 49% reuse rate ⚠️ — cache writes exceed reads on every run; cross-run caching is not occurring
• Error rate: 0/7 runs (0%) ✅
• Note: Very consistent runs — cost ranges from $0.49 to $0.54/run with nearly identical token counts, suggesting a deterministic input pattern with no cache warming across invocations.

Runs Without Token Data

6 runs were cancelled before completing and have no token data:

Run ID	Workflow	Conclusion
§24286063287	Smoke Claude	cancelled
§24286063276	Security Guard	cancelled
§24280111298	Security Guard	cancelled
§24280111303	Smoke Claude	cancelled
§24271299885	Security Guard	cancelled
§24271238328	Security Guard	cancelled

All 6 are cancelled runs — api-proxy data was not collected. No instrumentation changes needed.

---

Historical Trend

Comparing to the previous report (§1888 covering 2026-04-05 to 2026-04-10):

Workflow	Prev Avg Cost	Curr Avg Cost	Change	Prev Avg Turns	Curr Avg Turns	Change
Security Guard	$0.47	$0.22	✅ -53%	8.6	4.8	✅ -44%
Smoke Claude	$0.33	$0.19	✅ -42%	6.2	5.8	✅ -6%
Secret Digger (Claude)	$0.50	$0.51	➡️ +2%	3.0	3.0	➡️ stable

Security Guard and Smoke Claude both show significant cost improvements this period, likely due to improvements in agent efficiency or prompt changes. Secret Digger (Claude) remains flat — the cache inefficiency noted in prior reports persists unresolved.

---

Previous Report

§1888 — 2026-04-10 report

References:

• §24290378442 — this analyzer run
• §24290111812 — Security Guard highest cost run ($0.68, 11 turns)
• §24271368232 — Security Guard highest turn run (24 turns, 981K tokens)

Generated by Daily Claude Token Usage Analyzer · ● 466.4K · ◷