Kestrel Labs / Production / Usage
Usage analytics
Last 30 days
compared with the prior 30 days
Requests
1.62M
+29.2%
Tokens per request
2,393
+2.1%
Tokens
3.87B
+32.0%
Token anatomy
Input 3.09B
1,909 per request
+1.4%
Output 781.75M
484 per request
+5.0%
Output is 20.2% of tokens, up 0.5 pts on the prior 30 days.
What moved tokens
Prior
2.9B
More requests
+856.7M
Heavier requests
+81.0M
Now
3.9B
Tokens rose 937.8M. Request volume did most of it: +856.7M from more requests, +81.0M from heavier requests.
Daily volume
Average day
53.9K requests
Average day, tokens
128.94M
Peak day
Oct 12 · 76.2K
Days in window
30
01
Request and token traffic
1.62M requests, +29.2% on the prior 30 days. Claude Sonnet 4.5 leads with 42%.
Output anomaly
02
Token anatomy by model
Input and output tokens per request for each model, against the prior window. Select a model to focus the page on it.
← Input per request
Output →
Prior window · 3.09B input and 781.75M output tokens in total · Select a model to focus the page.
03
Model share shift
How request and token share moved between models, and how concentrated usage has become.
Prior → current
Effective models · requests
3.44
Usage is spread as if across 3.4 equally used models. Prior: 3.55.
Mix now
Top bar: now · thin bar: prior
Token weight
token share ÷ request share
Claude Sonnet 4.5
×1.24
GPT-4o
×0.93
DeepSeek V3
×0.90
Gemini 2.5 Flash
×0.68
Above ×1.00 a model uses more than its share of tokens per request. The tick marks parity.
04
Where usage comes from
Split consumption by calling workload or by API endpoint. Select a segment to filter the whole page to it.
Search and retrieval lost 0.9 pts of tokens share on the prior 30 days.
43.8%
Code review agent
22.8%
Support copilot
17.4%
Nightly summaries
16.0%
Search and retrieval
Workload
4 workloads · 1.62M requests · 3.87B tokens.
Select a segment to filter the page.
05
Patterns and anomalies
Findings computed from the usage data, each with the series it came from. Inspect one to scope the page to its window.
Daily requests
30 days
Low
High
Output anomaly
Average by weekday
Last 30 days
Mon
64.6K
Tue
59.9K
Wed
61.4K
Thu
60.6K
Fri
62.5K
Sat
36.7K
Sun
39.4K
Investigation queue
6 findings · Last 30 days
06
Usage records
Daily records per model and workload, the same grain as a provider usage export. They sum exactly to the totals above.
Oct 18
Claude Sonnet 4.5
Support copilot
Messages, streaming
7,431
11.54M
4.50M
2,158
$102.06
Oct 18
Claude Sonnet 4.5
Code review agent
Messages
3,995
15.51M
1.89M
4,353
$74.79
Oct 18
Claude Sonnet 4.5
Search and retrieval
Messages
2,202
4.50M
587.31K
2,312
$22.32
Oct 18
Claude Sonnet 4.5
Nightly summaries
Messages
4,007
10.01M
2.88M
3,217
$73.20
Oct 18
GPT-4o
Support copilot
Chat completions, streaming
1,516
1.99M
853.62K
1,879
$13.52
Oct 18
GPT-4o
Code review agent
Responses
480
1.58M
210.96K
3,732
$6.06
Oct 18
GPT-4o
Search and retrieval
Chat completions
2,093
3.63M
519.42K
1,981
$14.26
Oct 18
GPT-4o
Nightly summaries
Chat completions
2,028
4.30M
1.36M
2,787
$24.29
Oct 18
Gemini 2.5 Flash
Support copilot
streamGenerateContent
1,263
1.10M
402.26K
1,189
$1.34
Oct 18
Gemini 2.5 Flash
Code review agent
generateContent
137
297.47K
33.94K
2,419
$0.17
Oct 18
Gemini 2.5 Flash
Search and retrieval
generateContent
2,661
3.05M
373.51K
1,287
$1.85
Oct 18
Gemini 2.5 Flash
Nightly summaries
generateContent
3,741
5.24M
1.41M
1,779
$5.11
1–12 of 480 records · 1.62M requests · 3.87B tokens · $13,825 · matches the window totals