📊 Claude Token Usage Report2026-04-03

[github-actions]

Overview

Period: 2026-04-02T21:14Z to 2026-04-03T20:01Z (since previous report)
Runs analyzed: 4 of 5 Claude-engine runs (1 had no token data)
Total tokens: 1.22M across all workflows
Estimated total cost: $1.39

One new workflow gained token data this period: Security Guard — running 31 requests on a branch PR, it accounts for 81% of all tokens and 63% of total cost. Secret Digger (Claude) ran 3 times on main with consistent, efficient behavior.

---

Workflow Summary

Workflow	Runs	Requests	Total Tokens	Est. Cost	Cache Rate	I/O Ratio	Top Model
security-guard	1	31	985K	$0.87	99.1%	57:1	sonnet-4.6
secret-digger-claude	3	3 each	78–79K/run	$0.17/run	99.7%	38–53:1	sonnet-4.6

---

🔍 Optimization Opportunities

1. ℹ️ Security Guard — Haiku calls have zero cache hits — 10 haiku classification requests (all have cache_read_tokens = 0). At avg ~800 input tokens and total Haiku cost of $0.009 (1% of total), this is negligible today.

   • Recommendation: Low priority. Monitor if Haiku request count scales above 30/run before adding cache breakpoints.

2. ℹ️ Secret Digger — High cache write rate (50%) — Consistent with previous reports. Expected for a 3-request agent (first sonnet turn writes ~38K tokens; second turn reads them back). Cache reuse is healthy (1 write, 1 read per run).

   • Recommendation: No action needed. Pattern is structurally sound for a 2-sonnet-turn workflow.

3. ✅ Security Guard — Excellent cache efficiency — With 21 sonnet requests, 858K cache_read vs 103K cache_write gives a 8.3× cache reuse ratio. Cache writes are 10.7% of total, far below the 30% concern threshold.

---

Per-Workflow Details

security-guard

• Run: §23957630036 — perf/doc-maintainer-token-optimization branch (first run with token data)
• Requests: 31 total (10 haiku + 21 sonnet)
• Models: claude-haiku-4-5-20251001 (10 reqs), claude-sonnet-4-6 (21 reqs)
• Tokens: 985K total (8.2K input, 15.2K output, 858.8K cache_read, 103.4K cache_write)
• Cache hit rate: 99.1%
• Cache write rate: 10.7%
• I/O ratio: 57:1
• Avg latency: 8,600ms/request
• Estimated cost: $0.87

Cost breakdown (Sonnet):

• Cache writes: $0.39 (44%)
• Cache reads: $0.26 (29%)
• Output tokens: $0.22 (25%)
• Input tokens: $0.00 (<1%)
• Haiku (all): $0.01 (1%)

Request pattern: 1 haiku classification → first sonnet turn writes 37K cache tokens → 20 subsequent sonnet turns accumulate and read from cache. Each turn is ~1 token direct input, reading from 37K–82K cached context. Very efficient long-running conversation pattern.

---

secret-digger-claude

3 runs on main branch:

Run	Requests	Total Tokens	Cache Rate	I/O Ratio	Cost
§23959042026	3	78K	99.7%	53:1	$0.169
§23957139564	3	79K	99.7%	38:1	$0.174
§23955083235	3	79K	99.7%	40:1	$0.173

• Models: claude-haiku-4-5-20251001 (1 req/run), claude-sonnet-4-6 (2 reqs/run)
• Tokens per run: ~127 input, ~900 output, ~38.3K cache_read, ~39K cache_write
• Avg latency: 5,698–9,007ms/request
• Estimated cost: $0.172/run (avg)

Request pattern per run:

1. haiku — Input classification (123 tokens in, ~95 out, 0 cache, fast ~1.3s)
2. sonnet — First analysis turn (3 in, writes 38K cache, 6–27s)
3. sonnet — Second turn (1 in, reads 38K cache, adds 1K write, fast ~2.8s)

Workflows Without Token Data

• smoke-claude — 1 run (§23957630045, branch perf/doc-maintainer-token-optimization): artifact named agent not agent-artifacts; this run uses the Copilot engine, not the Claude engine with --enable-api-proxy. Excluded per scope rules.
• security-review — 0 runs this period
• smoke-claude (main) — 0 runs this period

---

Historical Trend

Previous report: 📊 Claude Token Usage Report2026-04-02 (issue #1629, 2026-04-02T21:14Z)

secret-digger-claude — Stable

Metric	Prev report (1 run)	Today (3 runs avg)	Change
Tokens/run	79K	78.4K	−1% ✅
Cost/run	$0.18	$0.172	−4% ✅
Cache hit rate	99.7%	99.7%	Stable ✅
Requests/run	3	3	Stable ✅

security-guard — New this period (first token data)

Security Guard ran for the first time with token tracking enabled. 31 requests across a long multi-turn session on a feature branch. Baseline established: 985K tokens, $0.87/run, 99.1% cache hit rate.

smoke-claude — No Claude-engine data

The smoke-claude run this period was the Copilot-engine variant; no Claude api-proxy data. Last Claude baseline: ~$0.24/run, ~203K tokens (from issue #1606).

---

Previous Report

📊 Claude Token Usage Report2026-04-02 — Issue #1629

References:

• §23957630036 — Security Guard (perf/doc-maintainer-token-optimization, largest run)
• §23959042026 — Secret Digger (Claude), most recent
• §23960188706 — This analyzer run

Generated by Daily Claude Token Usage Analyzer · ◷