fix(finops-framework): correct 1000x overstated cost and rescale the demo quotas

[J.K. (jk99k)]

Fixes #397. Two commits: correct the prices, then rescale the quotas that were calibrated against the wrong ones.

The prices

The Retail Prices API returns Foundry token meters with unitOfMeasure of 1K, so retailPrice is already a per-1000-token price. The notebook multiplies it by 1000 before writing it to PRICING_CL, but every consuming query already divides by 1000:

| extend InputCost  = PromptTokens     * InputTokensPrice
| extend OutputCost = CompletionTokens * OutputTokensPrice
| extend TotalCost  = (InputCost + OutputCost) / 1000

Each token ends up charged the price of a thousand tokens. For 1M input tokens on gpt-4.1-mini that's 400 USD reported instead of 0.40 USD.

The same KQL backs alert-suspend-sub and alert-activate-sub, so cost quotas trip about a thousand times too early and subscriptions get suspended almost as soon as they serve traffic.

The quotas

Correcting the prices on their own would leave the lab's central scenario unable to fire, so the second commit deals with that.

costQuota is 15/10/5 USD. Those values only work because costs are inflated: reaching the silver quota of 5 USD takes roughly 1,000 tokens today. With correct prices it takes 1,000,000, and the monthly tokenQuota of 1,000,000 is hit first, so no subscription can ever exceed its cost quota and the suspend/activate Logic App never runs.

Scaling the quotas by the same factor the prices moved, to 0.015/0.010/0.005, restores the previous behaviour exactly: whatever traffic tripped a quota before still trips it, and the dashboard shows the same relationship between CostQuota and TotalCost.

The variables cell now says these are demo-sized, so they aren't mistaken for realistic budgets and raised back to a magnitude that can't be reached.

Breaking change

No. Rows already in PRICING_CL keep their values, but consumers read with arg_max(TimeGenerated, *) by Model, so the corrected row wins on the next ingestion. Anyone who already tuned costQuota for their own traffic will want to rescale it the same way.

Type

Bugfix.

How to test

The premise needs no deployment:

curl -s --get 'https://prices.azure.com/api/retail/prices' \
  --data-urlencode "currencyCode='USD'" \
  --data-urlencode "\$filter=serviceName eq 'Foundry Models' and armRegionName eq 'swedencentral' and skuName eq 'gpt 4.1 mini Inp glbl'"

One item, retailPrice: 0.0004, unitOfMeasure: 1K. All six meters in the default models_config behave the same.

I also deployed the lab, ingested the corrected prices and ran the workbook's own KQL against the workspace. 1135 prompt + 1135 completion tokens on gpt-4.1 gives 0.01135; the same query on the old values gives 11.35. result.png here shows TotalCost: 10.072 for 2.27k tokens, which is that inflated magnitude.

What to check

The KQL is deliberately untouched. The workbook, dashboard and both alert rules already assume a per-1K price, so correcting the ingested value is enough and nothing else has to move.

The display-only table earlier in the notebook keeps its *1000; only its column header changed, to state the unit. That value is never ingested.

The markdown note that stated the incorrect assumption is corrected.

Other

Introduced in c329179 (#140). One file, +9/-6.

result.png still shows the old magnitudes and would want regenerating at some point, though that needs a run in your environment rather than mine.

[J.K. (jk99k)]

@microsoft-github-policy-service agree company="Microsoft"

[Copilot]

🟡 Changes recommended

Correct the pricing-unit note and remove its unconditional request-count claim.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR fixes a 1000× overstatement in Foundry token costs and rescales demo quotas accordingly.

Changes:

• Removes the erroneous * 1000 price transformation.
• Rescales demo quotas to 0.015/0.010/0.005.
• Updates pricing labels and documentation.

File summaries

File	Summary
labs/finops-framework/finops-framework.ipynb	Corrects pricing ingestion, quotas, display units, and related documentation.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.