Coverage report — example on synthetic data

Agent support-assistant · 18 Aug 2026 – 21 Aug 2026 · generated 30 Aug 2026

Observed spend
$2.05
$0.0103 per run
Runs
200
2.0% ended in error
Tokens processed
393 133
+ 36 660 read back from cache
Patterns observed
5
by semantic cluster

The tail is short

There are few enough distinct execution shapes to address all of them: there is no long tail for spend to disappear into, so there is nothing to choose between. This is the most favourable case — the 80/20 ratio does not apply here, because it assumes you have to select.

The 3 most expensive patterns — out of 5 in total — account for 90.5% of the spend, across 20.6% of the runs.

Where the spend concentrates

Cumulative share, patterns ordered from most to least expensive. The gap between the two lines is the point: a pattern can be half your runs and a rounding error on your bill.

share of spend share of runs
0%25%50%75%100% 80% threshold3 patterns 1 5 patterns, most expensive first pattern 1 of 5 · 50.7% of spend · 6.5% of runs · 13 runs · $1.04pattern 2 of 5 · 74.2% of spend · 8.5% of runs · 4 runs · $0.4821pattern 3 of 5 · 90.5% of spend · 20.6% of runs · 24 runs · $0.3342pattern 4 of 5 · 95.7% of spend · 45.7% of runs · 50 runs · $0.1074pattern 5 of 5 · 100.0% of spend · 100.0% of runs · 108 runs · $0.0871

Cost per pattern

The 5 patterns absorbing the most spend. These are not necessarily the most frequent ones: a rare, expensive pattern weighs as much as many frequent, cheap ones.

pat_bbf600a246b06686a4$1.04pat_bbf600a246b06686a4 · 13 runs · $0.0800 eachpat_3b1f9616a2d53b6c22$0.4821pat_3b1f9616a2d53b6c22 · 4 runs · $0.1205 eachpat_d197d790f677c8b6dc$0.3342pat_d197d790f677c8b6dc · 24 runs · $0.0139 eachpat_65a2c8d97ddd684b4e$0.1074pat_65a2c8d97ddd684b4e · 50 runs · $0.00215 eachpat_9a83b666c61ed528a7$0.0871pat_9a83b666c61ed528a7 · 108 runs · $0.00081 each

Addressable cost

The spend of the 3 patterns covering 80% of traffic.

$1.86 of $2.05 observed (90%)

This is not a projected saving. It is the spend a cheaper path could act on, and it becomes a saving only for those patterns that pass validation — run in shadow mode against live traffic, comparing the cheaper path with the current one and measuring how far they diverge. This report does not contain that measurement.

Cache keys you could widen

Input fields that varied while the result stayed the same. The cheapest saving here, since it needs no model.

No field could be shown inert. Every candidate varied together with request, so no pair isolates it. This is common on agents whose input is one blob of request text, and it is a property of the data rather than a failure of the measurement. 5 fields were shown to change the result.

Full distribution

Top twenty patterns by frequency.

#PatternRuns ShareCumulative CostAvg cost
1pat_9a83b666c61ed528a7f4a12d10854.0%54.0%$0.0871$0.00081
2pat_65a2c8d97ddd684b4e3d1d075025.0%79.0%$0.1074$0.00215
3pat_d197d790f677c8b6dcd98b832412.0%91.0%$0.3342$0.0139
4pat_bbf600a246b06686a4ac3567136.5%97.5%$1.04$0.0800
5pat_3b1f9616a2d53b6c222623da42.0%99.5%$0.4821$0.1205

Models and tools

Where the spend goes, by model invoked.

ModelCalls Cost
claude-haiku-4-5159$0.1952
claude-sonnet-524$0.3342
claude-opus-517$1.52

Most invoked tools: crm_lookup (122), ocr (50), erp_write (50), kb_search (37), ticket_update (13), sql_query (4), chart (4)

What this report does not tell you