Synthetic demonstration only. Invented usage, model names and rates; not a customer result or current pricing.
Scanner draft for human review. Signals are not confirmed waste; scenarios are not measured savings.
Analyzed: 1,100,000 input / 546,000 output tokens across 2 model(s) · $11.91 recorded spend
Export dates: 2026-10-01 → 2026-10-04 (dates available for 4 of 4 CSV rows).
Recorded project attribution
| Project/key label | Recorded spend |
|---|---|
| app-ingest | $0.19 |
| batch-pipeline | $11.71 |
| Unlabeled | $0.00 |
Three priorities to review
1. Spend concentrated in one model (high) — synthetic/review-model
98% of recorded spend is on synthetic/review-model ($11.71 of $11.91). Concentration is a review priority, not proof of waste.
2. Review response length (medium) — synthetic/review-model
Output is 37% of input plus output tokens; review whether response length is needed.
3. Reasoning tokens are part of output (medium) — synthetic/review-model
Reasoning is 71% of total output tokens; it is a subset, not an extra token bucket.
Details and next steps
1. Spend concentrated in one model — high
Where: synthetic/review-model
Observation: 98% of recorded spend is on synthetic/review-model ($11.71 of $11.91). Concentration is a review priority, not proof of waste.
Validation step: Benchmark suitable alternatives on the same tasks, including tokenization, cache eligibility, latency, quality and retries. Choose a model only after a measured comparison.
Conditional token-cost reduction: $9.37 for this export's workload; not measured savings.
Assumption: Same input, output (including reasoning), cache-read and cache-write counts on synthetic/alternative. Same cache eligibility, tokenization and accepted quality are unverified. Supplied flat rates apply across all requests; excludes fees, tools, tiers and retry changes. Compare modeled token costs, not the actual invoice total.
Modeled token costs: $11.71 at current supplied rates → $2.34 at candidate supplied rates. The recorded invoice can include other charges or different effective rates.
2. Review response length — medium
Where: synthetic/review-model
Observation: Output is 37% of input plus output tokens; review whether response length is needed.
Validation step: Test shorter responses on representative tasks; compare acceptance, retries and total cost before changing production limits.
Conditional token-cost reduction: $5.20 for this export's workload; not measured savings.
Assumption: 50% fewer total output tokens, including reasoning, with unchanged input, rates, retries and accepted task quality.
3. Reasoning tokens are part of output — medium
Where: synthetic/review-model
Observation: Reasoning is 71% of total output tokens; it is a subset, not an extra token bucket.
Validation step: Test a supported lower reasoning budget on representative tasks. Check quality and retries; hiding reasoning text does not stop its billing.
Conditional token-cost reduction: $3.70 for this export's workload; not measured savings.
Assumption: 50% fewer reasoning tokens, with unchanged visible output, input, rates, retries and accepted task quality. Reasoning is already included in output.
How to read these estimates
Scenarios overlap and are not added together. Output reductions, reasoning changes and model changes can affect the same tokens. No combined savings total or monthly extrapolation is provided.
Supplied prices are assumptions, not verified current quotes. Confirm their source, date, provider, cache accounting, tier and non-token fees before using a scenario in a customer recommendation. Token counts are held constant for comparison; alternative models may tokenize differently. Quality, retries and implementation feasibility require a measured trial.
Human teardown follow-up: reconcile the export to billing; validate the relevant price schedule; select feasible tests with the customer; report measured results separately from these scenarios.
What a human review would check
A human teardown would check the invoice and rate schedule, discuss what this workload must achieve, and agree a test before recommending a change. For this invented example, the first experiment would compare the current model with one candidate on representative tasks. Record accepted outputs, retries, latency and total cost for both. Keep the current model if the candidate fails the agreed quality bar.
This fixture uses a very small invented bill to make the arithmetic inspectable. It is not evidence that a $500 review would be worthwhile for a bill this small. We check whether your actual problem warrants paid work during the first look.
Request a free first look