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
-
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.
-
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.
-
✅ 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:
- 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:
haiku — Input classification (123 tokens in, ~95 out, 0 cache, fast ~1.3s)sonnet — First analysis turn (3 in, writes 38K cache, 6–27s)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:
Generated by Daily Claude Token Usage Analyzer · ◷
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
mainwith consistent, efficient behavior.Workflow Summary
🔍 Optimization Opportunities
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.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).
✅ 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
perf/doc-maintainer-token-optimizationbranch (first run with token data)Cost breakdown (Sonnet):
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
mainbranch:Request pattern per run:
haiku— Input classification (123 tokens in, ~95 out, 0 cache, fast ~1.3s)sonnet— First analysis turn (3 in, writes 38K cache, 6–27s)sonnet— Second turn (1 in, reads 38K cache, adds 1K write, fast ~2.8s)Workflows Without Token Data
perf/doc-maintainer-token-optimization): artifact namedagentnotagent-artifacts; this run uses the Copilot engine, not the Claude engine with--enable-api-proxy. Excluded per scope rules.Historical Trend
Previous report: 📊 Claude Token Usage Report2026-04-02 (issue #1629, 2026-04-02T21:14Z)
secret-digger-claude — 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: