Overview
Period: 2026-04-01T17:20Z to 2026-04-02T17:20Z (last 24 hours)
Runs analyzed: 5 of 21 Smoke Claude runs (16 additional runs not sampled due to time budget; pattern is consistent)
Total tokens (5 analyzed): ~1.01M across all workflows
Estimated total cost (5 runs): $1.19 | Estimated full-period cost (21 runs): ~$5.00
Token tracking data was available for all analyzed runs (all had agent-artifacts with token-usage.jsonl). No previous Claude token usage report issues found — this is the first report. Historical trend analysis will be available in future reports.
Note: Only Smoke Claude had active runs this period. Security Guard had 1 skipped run (no token data). All other Claude-adjacent workflows (secret-digger-claude, security-review) did not run.
Workflow Summary
| Workflow |
Runs |
Runs Analyzed |
Total Tokens |
Cost (analyzed) |
Est. Cost (all) |
Cache Rate |
I/O Ratio |
Top Model |
| smoke-claude |
21 |
5 |
~1.01M |
$1.19 |
~$5.00 |
99.4–99.7% |
109–168:1 |
sonnet-4.6 |
| security-guard |
1 |
0 |
— |
— |
— |
— |
— |
— |
🔍 Optimization Opportunities
-
⚠️ High I/O ratio on smoke-claude — All 5 analyzed runs show ratios of 109–168:1 (threshold: >100:1)
- The Sonnet model reads ~772K cached tokens vs only ~31 direct input tokens across all runs
- Root cause: The large context/system prompt is served almost entirely from cache — this is expected behavior given 99.5%+ cache hit rates, not a true inefficiency
- Recommendation: No action needed; the high ratio is a consequence of excellent caching. Monitor if ratio increases further, which could indicate prompt expansion.
-
ℹ️ Haiku requests have zero cache hits — 10 Haiku requests across 5 runs, all with cache_read_tokens = 0
- These are small classification/routing calls (~400 input tokens, ~50 output tokens)
- Since Haiku is already very cheap ($0.80/M input) and the total Haiku cost is ~$0.005 (0.4% of total), adding prompt caching here offers minimal savings
- Recommendation: Low priority. If Haiku usage scales significantly (>100 requests/run), consider adding cache breakpoints.
-
💡 Cache writes are the largest cost driver — cache_write_tokens account for $0.87 of $1.19 total (73%) across the 5 analyzed runs
- Sonnet writes ~40–58K cache tokens per run, then reads ~138–193K on subsequent turns
- This is healthy cache reuse behavior (read/write ratio ≈ 3–4x)
- Recommendation: No action needed currently. If cache_write_rate increases above 30% consistently, review whether conversation turns are long enough to justify the write cost.
Per-Workflow Details
smoke-claude
5 representative runs analyzed (21 total this period):
| Run |
Branch |
Requests |
Total Tokens |
Cache Rate |
I/O Ratio |
Cost |
| §23900672204 |
main |
7 |
254K |
99.7% |
119:1 |
$0.30 |
| §23911242765 |
feat/claude-token-usage-analyzer |
6 |
198K |
99.4% |
111:1 |
$0.28 |
| §23882121359 |
fix/copilot-session-state-chroot-mount |
6 |
194K |
99.5% |
165:1 |
$0.22 |
| §23882382764 |
fix/copilot-session-state-chroot-mount |
6 |
194K |
99.5% |
168:1 |
$0.21 |
| §23908898243 |
fix/openai-cache-token-tracking |
6 |
175K |
99.5% |
109:1 |
$0.18 |
Model distribution (across 5 runs, 31 total requests):
| Model |
Requests |
Input |
Output |
Cache Read |
Cache Write |
Avg Latency |
| claude-sonnet-4-6 |
21 (68%) |
31 |
5,535 |
772,360 |
232,010 |
6,117ms |
| claude-haiku-4-5-20251001 |
10 (32%) |
3,600 |
454 |
0 |
0 |
775ms |
Token breakdown (5 runs):
- Input tokens: ~4K (0.4%)
- Output tokens: ~6K (0.6%)
- Cache read tokens: ~772K (76.2%) ← primary I/O channel
- Cache write tokens: ~232K (22.9%)
- Grand total: ~1.01M tokens
Cost breakdown (5 runs):
- Sonnet cache writes: $0.87 (73%)
- Sonnet cache reads: $0.23 (20%)
- Sonnet output: $0.08 (7%)
- Haiku (all): $0.005 (<1%)
Typical request pattern per run:
haiku — Tool/task classification (~400 input tokens, fast 532ms)sonnet — First turn, writes cache (~40K tokens, 6.6s)haiku — Secondary classification (~400 input tokens, fast 531ms)sonnet — Turn 2, reads+extends cache (4.3s)sonnet — Turn 3, reads+extends cache (11.3s)sonnet — Turn 4, small response from cache (2.3s)
Workflows Without Token Data
The following workflows either didn't run, were skipped, or don't use --enable-api-proxy:
- security-guard — 1 run (skipped; no token data)
- secret-digger-claude — 0 runs this period
- security-review — 0 runs this period
Copilot-engine workflows (smoke-copilot, smoke-chroot, smoke-services, build-test, smoke-codex) and Codex-engine workflows are excluded from this analysis — they are covered by the Copilot Token Usage Analyzer.
Historical Trend
This is the first Claude token usage report. No previous reports found under the claude-token-usage-report label. Historical trends will be available in future daily reports.
Baseline established (2026-04-02):
- Per-run cost: ~$0.24
- Per-run tokens: ~203K
- Cache hit rate: ~99.5%
References:
Generated by Daily Claude Token Usage Analyzer · ◷
Overview
Period: 2026-04-01T17:20Z to 2026-04-02T17:20Z (last 24 hours)
Runs analyzed: 5 of 21 Smoke Claude runs (16 additional runs not sampled due to time budget; pattern is consistent)
Total tokens (5 analyzed): ~1.01M across all workflows
Estimated total cost (5 runs): $1.19 | Estimated full-period cost (21 runs): ~$5.00
Token tracking data was available for all analyzed runs (all had
agent-artifactswithtoken-usage.jsonl). No previous Claude token usage report issues found — this is the first report. Historical trend analysis will be available in future reports.Workflow Summary
🔍 Optimization Opportunities
smoke-claude— All 5 analyzed runs show ratios of 109–168:1 (threshold: >100:1)ℹ️ Haiku requests have zero cache hits — 10 Haiku requests across 5 runs, all with
cache_read_tokens = 0💡 Cache writes are the largest cost driver —
cache_write_tokensaccount for $0.87 of $1.19 total (73%) across the 5 analyzed runsPer-Workflow Details
smoke-claude
5 representative runs analyzed (21 total this period):
mainfeat/claude-token-usage-analyzerfix/copilot-session-state-chroot-mountfix/copilot-session-state-chroot-mountfix/openai-cache-token-trackingModel distribution (across 5 runs, 31 total requests):
Token breakdown (5 runs):
Cost breakdown (5 runs):
Typical request pattern per run:
haiku— Tool/task classification (~400 input tokens, fast 532ms)sonnet— First turn, writes cache (~40K tokens, 6.6s)haiku— Secondary classification (~400 input tokens, fast 531ms)sonnet— Turn 2, reads+extends cache (4.3s)sonnet— Turn 3, reads+extends cache (11.3s)sonnet— Turn 4, small response from cache (2.3s)Workflows Without Token Data
The following workflows either didn't run, were skipped, or don't use
--enable-api-proxy:Historical Trend
This is the first Claude token usage report. No previous reports found under the
claude-token-usage-reportlabel. Historical trends will be available in future daily reports.Baseline established (2026-04-02):
References: