Kestrel Labs / Production
Overview
Spend · selected range
$13,825
+36.4%
vs prior 30 days
$461 per day
Over the last 30 days, spend grew 36.4%, outpacing request volume (+29.2%). Claude Sonnet 4.5 accounts for 87% of that change.
API requests
1.6M
+29.2%
vs prior 30 days
53.9K per day on average
Total tokens
3.87B
+32.0%
vs prior 30 days
20% output · 2.39K per request
Success rate
99.50%
-0.06 pts
vs prior 30 days
8.1K failed requests
P95 latency
2.75s
+3.8%
vs prior 30 days
95th percentile, request-weighted
Cost per 1K requests
$8.55
+5.5%
vs prior 30 days
Blended across selected models
01
What moved spend
Spend rose $3,687 over the last 30 days. Request volume explains 80% of the change, so growth is coming from usage rather than price or model choice.
Hover or focus an effect to see which models drive it. The three effects add up exactly to the change in spend.
02
Needs attention
4 to act on
4 to watch
Last 30 days
Budget · workspace
Act
Projected to exceed the monthly budget by $657
At $512 a day (weekday-adjusted), month-end lands at $15,657, crossing $15,000 on Oct 30.
$14,486–$16,829 likely range
Cost anomaly · Oct 12
Act
Claude Sonnet 4.5 spend ran 85% above its 7-day average
Answers grew from 563 to 970 output tokens per request. Volume was normal, so a prompt or template change is the likely cause.
+$326 that day
Cost anomaly · Oct 9
Act
Claude Sonnet 4.5 spend ran 36% above its 7-day average
Request volume stepped up while tokens per request held steady, consistent with a launch or new traffic source.
+$132 that day
Cost anomaly · Sep 25
Act
Claude Sonnet 4.5 spend ran 55% above its 7-day average
Request volume stepped up while tokens per request held steady, consistent with a launch or new traffic source.
+$154 that day
Reliability · Oct 15
Watch
DeepSeek V3 errors reached 1.5%
3.0× the model's normal error rate. Failed calls are retried, so this also inflates request counts.
Retries add cost
3 more in this period. Narrow the range or models to see them.
03
Usage and spend
Requests up 29.2% vs prior 30 days, accelerating into the latest stretch. Busiest day: Oct 12 · 76.2K
Claude Sonnet 4.5
GPT-4o
Gemini 2.5 Flash
DeepSeek V3
Daily spend (right scale)
Prior period
04
Month-end forecast
Projected month-end
$15,657
80% range $14,486–$16,829
Spent so far
$8,864
Day 18 of 31
Daily allowance
$472
To stay within $15,000, vs $512 run rate
Budget runway
Oct 30
Crosses budget with 1 day left
Actual daily spend
Forecast with 80% band
Cumulative spend
05
Savings opportunities
Identified per month
$3,217
Add an opportunity to the forecast to see its effect on month-end spend.
Claude Sonnet 4.5
Gemini 2.5 Flash
Route short support replies to Gemini 2.5 Flash
35% of Support copilot traffic on Claude Sonnet 4.5 is short-form. Flash serves a request for 7% of the cost.
Assumes 35% of Sonnet support requests qualify and quality holds on your eval set.
$1,339
/mo
Needs evaluation
Claude Sonnet 4.5
Cache repository context in the code review agent
About 55% of each Code review agent prompt repeats the same repository context. Cached input is billed at 10% of the standard rate.
Net of cache-write overhead (80% of the gross saving).
$939
/mo
Code change
GPT-4o
Move nightly summaries to batch processing
Nightly summaries are not latency-sensitive. Batch requests on OpenAI and Anthropic are billed at 50%.
Gemini and DeepSeek traffic excluded.
$601
/mo
Config change
GPT-4o
Gemini 2.5 Flash
Re-rank search results with Gemini 2.5 Flash
50% of Search and retrieval traffic on GPT-4o is re-ranking short passages. Flash serves a request for 13% of the cost.
Assumes half of GPT-4o search requests qualify and ranking quality holds on your eval set.
$338
/mo
Needs evaluation
Claude Sonnet 4.5
Restore the output token cap on Claude Sonnet 4.5
Answers grew from 563 to 970 output tokens per request after the Oct 12 template change.
Recovers the excess output tokens observed during the regression, if it recurs.
$203
/mo
Config change
Prevents a repeat. Not in the forecast.
Estimates use the last 30 days of usage at list prices. Simulating an opportunity changes only the forecast on this page.
06
Model mix
Claude Sonnet 4.5 carries 42% of requests but 79% of spend, 1.9× its traffic weight. DeepSeek V3 serves 1K requests for $0.98.
Last 30 days · spend and shares across all 4 models. Select a model to focus the whole overview on it.
07
Reliability
All selected models operational
Success rate
99.50%
P95 latency
2.75s
Failed requests
8.1K
Daily error rate
Peak 1.00%
Detected incidents · 4
Oct 15
DeepSeek V3 error spike
1.46%
Oct 14
Gemini 2.5 Flash latency degradation
3.35s
Oct 9
GPT-4o error spike
4.58%
08
Spend by workload
Last 30 days · attributed from request tags. Select a workload to see its model costs.
09
Request rhythm
Peak weekday window is 13:00–17:00 UTC. Only 8% of requests lands overnight, leaving room for batch jobs.
Last 30 days · hover a cell for detail
Low
High
10
Activity
Incidents and budget crossings are detected from usage. Configuration and key changes come from the workspace audit trail.
Oct 18
14:22 UTC
Security
API key prod-ingest rotated
Previous key revoked after a 24-hour overlap window.
Oct 16
10:04 UTC
Config
Fallback route added
GPT-4o requests now fail over to Claude Sonnet 4.5 after two retries.