THE FINANCE READ
AI COST LENS · FREE TOOL
Is the cheaper AI
actually costing you more?
Compare two tools or plans, including the time you spend checking and fixing their work. Or start with a bill to see the costs you can account for.
No account or connection needed. Your files are read in this browser.
A lower subscription price is only part of the picture. Extra review time and fewer usable outputs can change which option costs less.
Read the guide
PROVIDER REPORTED
OPENAI BILL REVIEW
What can this bill actually prove?
Saved dashboard exports
Total cost and usage mix are real evidence. Unsupported allocations stay blank.
YOUR AI TOOLS
Subscriptions and extra usage, together
Included usage stays inside the subscription. Only separately charged usage is added.
| Tool / charge | Owner | Source reference | Period amount |
|---|
BILL TREND
How did the bill change?
BUDGET CHECK
How does this bill compare with the plan?
USAGE MIX
Where the requests and tokens went
These are observed usage measures. They are not billed dollars by model.
| Model | Requests | Input | Output | Cache share | Billed cost |
|---|
Billed cost by project
Cost export attribution. A project total is not a billed model total, even if the usage file shows one model for that project.
| Project | Billed cost | Share of bill |
|---|
THE JUDGMENT
What to act on, and what to leave alone
This separates what the bill proves from what still needs testing.
NEXT EVIDENCE
FINANCE FIRST AI SPEND REVIEW
What did one ready result really cost?
- Cost finding
- Before switching
EXPLORE TWO EXAMPLES
STARTUP UNIT ECONOMICS
Would the margin hold as usage grows?
Try your own sale price per ready result. Both routes keep their example ready rates; infrastructure can stay fixed while model and review costs scale.
Enter revenue to see a modeled gross margin at 1×, 2×, 5×, and 10× volume.
Illustrative scenario only. Revenue applies to ready results; the entered discount applies to both model bills; review effort uses the entered share at every volume. Quality, prices, fixed costs, demand, and capacity may change. One-time change cost is excluded from gross margin.
THE RECEIPT
What one ready result costs
Model usage is one part of the cost. The evidence label travels with the numbers.
TOKEN-TO-COST CHECK
Do the tokens support this amount?
THE FINANCE READ
THE WORK
How the two routes compare
Total spend can mislead when volume or quality changes. This view uses the same ready-to-use outcome on both sides.
PLAN CHECK
Did the current route land where expected?
The plan, the actual result, and the proposed change stay separate. A lower forecast is not the same as money already saved.
Same workload and period
| Measure | Plan | Actual | Variance (actual − plan) |
|---|
WHAT THE COMPARISON FOUND
What the comparison says
The next step follows the cost per ready result and the checks behind it.
THE DECISION
RECURRING COST BOUNDARY
MODEL + SHARED INFRASTRUCTURE + HUMAN REVIEWCOST ANATOMY
The bill is bigger than tokens.
A cheaper rate can disappear after retries, review, or infrastructure. The breakdown keeps each part visible instead of burying everything in one total.
WORKLOAD BEHAVIOR
Where the economics changed
These measures cover all provider requests, including retries. Cost per ready result uses only outputs that cleared the quality rule.
FINANCIALLY RANKED FINDINGS
OVERLAP REMOVED FROM THE HEADLINEOPPORTUNITIES
Where is the credible opportunity?
Each finding shows its evidence basis, financial confidence, overlap, and next step. Candidate amounts are never added together and called savings.
REQUEST-LEVEL REVIEW
Find the calls behind the pattern
Add a CSV or JSON request log, including a LiteLLM spend-log export, to break down spend, calculate an evidence-gated run rate, and detect duplicate records, failed and repeated calls, retry loops, repeated uncached prefixes, oversized inputs, excessive outputs, reasoning intensity, and repeated tool use. Gateway spend remains separate from provider billing evidence. Your source file is parsed in this browser and is not uploaded.
Rows with invalid values or the wrong number of CSV fields are excluded from a scoped review. Download the issue list, correct the source, and analyze again for a whole-log comparison.
Use one row per AI request or trace observation. For LiteLLM, export all pages of individual spend logs; a single paginated response is incomplete. The CLI canonical CSV is a separate period-level billing format and is not interchangeable with this request log.
Budget and allocation settings (optional)
Add compute, data, network, tooling, pipeline, and human costs
Use totals for the same period and currency as the request log. Enter only costs not already included in provider request cost or another category. Blank means unknown. Enter zero only when you confirmed there was no cost. AI Cost Lens shows a fully loaded total only when every request is priced and every category is supplied.
Costs and findings cover valid rows only. Bill comparison, run rate, and whole-log claims are unavailable until you correct and reimport the excluded rows.
OBSERVED REQUEST REVIEW
What the imported calls support
Spend context and operating signals
SPEND CONTEXT
What shaped the imported cost?
Local snapshot · not live monitoring
OPERATING SIGNALS
What could be driving the cost?
Coverage is shown with every metric
PERIOD VARIANCE
Why did the recent window change?
Equal consecutive UTC windows
Customer cost to serve
COST TO SERVE
Which customers carry the AI cost?
Join customer IDs from this request log to revenue covering the log dates in the same currency. Extra days in the revenue period make the cost share a lower bound. Optionally spread entered operating costs, with missing and unattributed amounts shown.
Request rows and filters
LOCAL SPEND EXPLORER
Trace the finding back to its rows
| Event | Time | Provider · model | Project · workload | Status | Input | Output | Cache | Selected cost |
|---|
Import coverage and limitations
OBSERVED SHAPE · PROPOSED PRICE
SIMULATION IS NOT VERIFICATIONSIMULATE
What if the workload changed?
Reprice the observed token shape under one proposed model or operating change. Quality, latency, tools, and model behavior remain unproven until you test them.
TEST FIRST
WHAT STAYED FIXED
SAME WORK · SAME RULE · BLIND WHEN POSSIBLE
QUALITY BEFORE ROUTE CHANGEVERIFY
Did the cheaper route still do the job?
Price can nominate a candidate. A comparable output sample determines whether it belongs in the decision.
THE TEST
Use one decision rule for both routes
Review model names only after scoring when practical.
- 01Choose the same tasks
Use representative work from one bounded workload.
- 02Hide the route
Blind model names and randomize order when practical.
- 03Apply the same rule
Mark ready, needs correction, or needs escalation and record review time.
- 04Return to the decision
Cost, quality, policy, and evidence must pass together.
PAIRED OUTPUT CHECK
Test the same cases locally
Import one baseline row and one candidate row for every case. AI Cost Lens can check valid JSON, required or forbidden terms, required JSON fields, an exact label, and a maximum length. Separate multiple terms or JSON fields with a semicolon (;), as shown in the separate illustrative example. Human outcomes and review time remain separate evidence.
Score the outputs here
AI Cost Lens will hide route names, randomize A/B order separately for every case, and clear the rendered outputs after the last score.
Score this pair without route names
Apply the same ready-result definition to A and B. The route is revealed only in the private calculation after scoring.
Already scored elsewhere? Keep the outcomes in the CSV, then describe how that review was run.
Every case must use the same deterministic rules on both routes. Arbitrary regular expressions and executable schemas are deliberately not run.
QUALITY EVIDENCE · NOT A SAVINGS CLAIM
What the paired cases support
| Case | Baseline checks | Candidate checks | Baseline outcome | Candidate outcome |
|---|
DECISION RECORD
NO PROVIDER CREDENTIALS REQUIREDEVIDENCE CHECK
What can we actually prove?
Every conclusion should show where the cost came from, what counted as a good result, what was reviewed, and what is still missing.
NEXT RECORDS TO COLLECT
Make the next test count
The workspace archive includes completed sections and customer IDs when supplied. Keep it private. “Open saved review” accepts the decision record separately.
ACCOUNTING BOUNDARY
POST-CHANGE RECONCILIATION
MODELED → OBSERVED → REALIZEDACTUALS
Did the savings appear after the change?
Enter the post-change period from your records. Compare recurring costs at the same ready-result volume, then show any one-time implementation cost separately. Amounts stay in the review currency.
SAVINGS LEDGER
One-time implementation cost is deducted in this period only. The recurring difference excludes it; subsequent periods need their own actuals.