📊 Claude Token Usage Report2026-05-02

[github-actions]

Overview

Period: 2026-05-01T21:48Z to 2026-05-02T19:00Z (~21 hours)
Runs analyzed: 48 (22 had token data; 26 missing — early exits or lean runs)
Total tokens: 5,174K (effective: 1,697K with caching applied)
Estimated total cost: $3.39
Total Actions minutes: 162 min

⚠️ 26 of 48 runs had no token data (18× Security Guard, 8× Smoke Claude) — these completed before or without the agent generating tokens.

Workflow Summary

Workflow	Runs	Runs w/ Tokens	Total Tokens	Avg Tokens	Total Cost	Avg Cost	Avg Turns
Security Guard	24	6	2,531K	422K	$2.30	$0.38	10.3
Smoke Claude	24	16	2,643K	165K	$1.09	$0.07	6.7

🔍 Optimization Opportunities

1. Security Guard — avg $0.38/run, avg 422K tokens/run, 10.3 turns/run

   • High token count (>100K threshold) — review if the system prompt or tool surface can be reduced
   • High turn count (>10 avg) — consider pre-computing deterministic checks as workflow steps before invoking Claude
   • 75% of runs missing token data (18/24) — these may be very short runs that exit early (no cost) or have instrumentation gaps; worth confirming all runs use the api-proxy
   • 5/24 runs had errors (21% error rate)

2. Smoke Claude — high error rate: 13/24 runs errored (54%)

   • Error rate well above the 30% flag threshold — investigate reliability issues in the smoke test workflow
   • Cache efficiency is 77% (reads vs. writes) — healthy but could improve further with longer conversation turns

Cache Analysis

Both workflows use Anthropic prompt caching effectively:

Workflow	Cache Reads	Cache Writes	Cache Efficiency
Security Guard	2,541K	406K	86.2% ✅
Smoke Claude	2,022K	600K	77.1% ✅

Cache efficiency is healthy for both workflows — the system prompt is being cached and reused across turns.

Per-Workflow Details

Security Guard

• Runs: 24 total (6 with token data, 18 with no token data)
• Models: claude-haiku-4-5-20251001, claude-sonnet-4-6
• Total tokens: 2,531K (avg 422K/run on runs with data)
• Effective tokens: 783K (heavily cached)
• Estimated cost: $2.30 total (avg $0.38/run with data)
• Turns: 62 total (avg 10.3/run)
• Cache: 2,541K reads / 406K writes — 86.2% efficiency
• Actions minutes: 62 min
• Error rate: 5/24 runs (21%)

Recent runs (latest 5):

• 2026-05-02T17:03Z — 412K tok, $0.41, 9 turns ✅
• 2026-05-02T17:19Z — 467K tok, $0.41, 11 turns ❌
• 2026-05-02T17:23Z — 292K tok, $0.26, 9 turns ✅
• 2026-05-02T18:05Z — no token data, $0.00 ✅
• 2026-05-02T19:00Z — no token data, $0.00 ✅

Smoke Claude

• Runs: 24 total (16 with token data, 8 with no token data)
• Model: claude-haiku-4-5-20251001 only
• Total tokens: 2,643K (avg 165K/run on runs with data)
• Effective tokens: 914K
• Estimated cost: $1.09 total (avg $0.07/run with data)
• Turns: 107 total (avg 6.7/run)
• Cache: 2,022K reads / 600K writes — 77.1% efficiency
• Actions minutes: 100 min
• Error rate: 13/24 runs (54%) ⚠️

Recent runs (latest 5):

• 2026-05-02T17:03Z — 168K tok, $0.07, 7 turns ✅
• 2026-05-02T17:19Z — 126K tok, $0.07, 6 turns ✅
• 2026-05-02T17:23Z — 169K tok, $0.05, 8 turns ✅
• 2026-05-02T18:05Z — 125K tok, $0.07, 6 turns ✅
• 2026-05-02T19:00Z — 126K tok, $0.07, 7 turns ✅

Runs Without Token Data

26 runs had token_usage = null or 0. These may have exited early (before the agent started) or completed without invoking the LLM. Workflows affected:

• Security Guard: 18 runs with no token data (75% of its runs)
• Smoke Claude: 8 runs with no token data (33% of its runs)

If these are unexpected, verify that the api-proxy sidecar is enabled for all agentic runs of these workflows.

Historical Trend

Date	Runs	Total Tokens	Total Cost	Error Rate
2026-04-29	—	—	—	—
2026-04-30	—	—	—	—
2026-05-01	42	7,847K	$6.14	—
2026-05-02	48	5,174K	$3.39	37.5%

Cost decreased ~45% from yesterday ($6.14 → $3.39), possibly due to fewer high-cost Security Guard runs completing with token data today.

Previous Report

📊 Claude Token Usage Report 2026-05-01 — Issue #2373

References:

• §25260610822 — this report's workflow run

Generated by Daily Claude Token Usage Analyzer · ● 426.3K · ◷