Measure cache use
See fresh input, cache writes, cache reads, and output tokens.
LLM PROMPT-CACHE ANALYTICS
Cut LLM costs with better prompt caching.
CacheEconomics analyzes LiteLLM usage metadata, shows where repeated input misses the cache, and recommends the first fix to test.
- Works with LiteLLM today
- Prompt and response text rejected
- Open-source analyzer runs offline
RESEARCH AGENTCache overview
Receiving data
Requests analyzed286Accepted usage events
Cached input16.7%Share of input tokens
Recommended fixes4Ranked by likely impact
Savings estimateAdd pricingUse your contract rates
Cached share of input tokens7 days
9%Mon
11%Tue
10%Wed
14%Thu
13%Fri
16%Sat
17%Sun
Top actionTTL-1
Keep stable instructions together so more requests can reuse them.
BASED ON REQUEST STRUCTURE
Live LiteLLM callback
Prompt and response text rejected
Multi-tenant access controls
Open-source core
WHAT CACHEECONOMICS SHOWS
Where repeated input is missing the cache.
System prompts, tool definitions, documents, and conversation history often repeat. See how much is reused and what still pays the standard input rate.
Find missed reuse
Spot repeated input that expires or changes before it can be reused.
Estimate financial impact
Apply your pricing to measured traffic and see how each estimate was calculated.
Monitor data delivery
Track request volume, latency, errors, analysis jobs, and ingestion lag.
SAVINGS CALCULATOR
Estimate your prompt-caching savings.
Use three numbers from your LLM bill and traffic pattern.
DASHBOARD TOUR
Watch the dashboard find a cache issue.
These short tours use example data, so you can see the workflow without exposing customer traffic.
EXAMPLE DATA · DASHBOARD TOUR00:13
HOW IT WORKS
From LLM requests to practical cache fixes.
-
01
Collect usage metadata
The collector records token counts, timing, outcomes, and private identifiers for repeated prompt sections.
-
02
Block prompt text
Every event is checked before storage. Prompt text, response text, and unknown fields are rejected.
-
03
Measure cache performance
The analyzer compares fresh input, cache writes, cache reads, request timing, and repeated sections.
-
04
Rank the next fixes
The dashboard explains the observed problem, the change to test, and any risk to response quality.
PRIVACY BOUNDARY
Your prompt and response text stay in your environment.
The collector sends token counts, timing, model names, outcomes, and one-way identifiers. It rejects prompt and response text.
Read the security model
STAYS WITH YOU
Prompt text
Response text
Secret hashing key
SENT TO THE DASHBOARD
Input and output token counts
Latency and request status
Private section identifiers
OPEN ENGINEERING
Review how it was built.
The tests, system design, security decisions, and remaining production work are public.
Automated testsPrivacy checks, regression tests, and container scans run on every change.↗
System designSee how authentication, organizations, data ingestion, and analysis fit together.↗
Production readinessSee what has been tested, deployed, and left on the readiness checklist.↗
Source code and licenseRead the Apache-2.0 code or run the analyzer locally.↗
FAQ
Common questions.
What can I connect today?
You can connect a LiteLLM proxy callback, upload existing LiteLLM JSONL logs, or upload a normalized CacheEconomics trace.
Does CacheEconomics read or store my prompts?
No. The collector sends usage metadata and one-way identifiers. Prompt text and response text are rejected before they can enter the delivery queue or hosted API.
How does the savings calculator work?
It multiplies monthly input-token spend by the share of reusable input and the cache-read price reduction. The estimate excludes output tokens, cache-write charges, negotiated rates, and implementation cost.
Why can a dashboard estimate be unavailable?
The dashboard needs enough complete request data and applicable pricing or invoice information. When either is missing, it explains what is needed instead of showing an unsupported dollar figure.
Can a team share one workspace?
Yes. Organizations keep each team’s data separate, and roles control who can view results, manage sources, run analyses, and administer members.
What data is shown in the public demo?
The public demo uses generated test traffic. It lets visitors explore the product without exposing customer prompts, usage, invoices, or performance data.
READY TO ANALYZE YOUR TRAFFIC?
Find your biggest cache opportunity.
Connect LiteLLM or upload a usage file. See where cached input is being missed and what to change.