Where did all your tokens go? Managing AI spend with the Coralogix AI Centre
Where did all your tokens actually go?
In this video, Annie uses the Coralogix AI Center and Olly to find the three changes that cut token spend the most, across the AI apps you build and the coding agents your team uses every day.
You'll see why re-reading context makes up most of a token bill (63% of spend across 25,000 coding agent sessions we studied), how a few users and sessions can drive almost half of your usage, and how prompt caching can cut input token costs on the repeated part of your prompts.
Chapters
00:00 - Where did all your tokens go?
00:06 - 1. Send less context
00:45 - Why re-reading context dominates the bill
01:05 - Asking Olly why conversations run long
01:42 - The highest-impact fixes
02:21 - How much Olly says you could save
02:48 - Tracing a span with OpenTelemetry
03:11 - 2. Find the few sessions and users driving spend
03:29 - Three users, 46% of Claude Code tokens
04:20 - Route routine work away from Opus
04:52 - Tracking your cache hit rate
05:09 - 3. Use prompt caching
05:46 - What to cache, and what never to cache
06:26 - Putting it together
Connect with Annie on LinkedIn: https://www.linkedin.com/in/annie-freeman-speaker-engineer-advocate
Learn more about Coralogix: https://coralogix.com/about/
Subscribe to our channel for more insights into observability and AI.
#tokens #promptcaching #aiobservability #claudecode #codingagents #opentelemetry #llm