1,000x Overcharge Bug on "Qwen38 27b SFT Gl 1M Tokens" (44K tokens billed as 44M tokens)

Akihisa Hirosawa 0 Reputation points
2026-10-29T03:34:32.78+00:00

Issue Summary

I ran a small Supervised Fine-Tuning (SFT) job on qwen3.8-27b (Global training tier) in eastus2.

Although the Microsoft Foundry portal explicitly confirms that only 44,000 tokens (0.044M tokens) were billed, Azure Cost Management recorded the usage quantity as 44 (44M tokens) against a 1M token meter, resulting in an exact 1,000x overcharge ($132.00 USD instead of $0.132 USD).

I opened a Billing Support request (SR #2610200010000024), but the vendor support team insists that because 44 × $3.00 = $132.00, the invoice math is "valid," and they refuse to investigate the metering pipeline bug unless I purchase a paid Technical Support plan.

Expected vs. Actual Billing Comparison

  • Model & Tier: qwen3.8-27b (SFT Global)
  • Meter Name: Qwen38 27b SFT Gl 1M Tokens
  • Meter ID: 59a176f8-721a-5e7d-954a-250a4e3a5e3d
  • Unit of Measure: 1M (1 Million Tokens)
  • Official Unit Price: $3.00 USD per 1M tokens
  • Tokens Billed in Foundry UI (Expected): 44,000 tokens (0.044M tokens)
  • Quantity Recorded in Azure Billing (Actual): 44 (44,000,000 tokens / 44M tokens) — 1,000x unit conversion bug
  • Expected Charge: $0.132 USD (¥20.80 JPY)
  • Actual Billed Charge: $132.00 USD (¥20,796.60 JPY / displayed as ¥20,797)
  • Overcharged Amount: +$131.87 USD (+¥20,775.80 JPY)

Evidence & Primary Data

1. Microsoft Foundry Portal (Screenshot Attached)

  • Job Status: Completed in 14m (actual training runtime: ~28 seconds across 2 epochs / 2 steps)
  • Base Model: qwen3.8-27b (Supervised / Global)
  • Training tokens billed: 44,000

2. Official Azure Retail Prices API (eastus2) The new meter (effective October 1, 2026) is priced per 1M tokens with tierMinimumUnits set to 0.0:

  • meterName: Qwen38 27b SFT Gl 1M Tokens
  • meterId: 59a176f8-721a-5e7d-954a-250a4e3a5e3d
  • skuId: DZH318Z0SDFC/01PV
  • armRegionName: eastus2
  • unitOfMeasure: 1M
  • unitPrice: 3.0 USD
  • tierMinimumUnits: 0.0
  • effectiveStartDate: 2026-10-01T00:00:00Z

3. Actual Usage Record from Microsoft.Consumption/usageDetails API Unlike older fine-tuning meters (such as Qwen3 32B FT Tokens) which are billed per 1K tokens, Qwen38 27b SFT Gl 1M Tokens is billed per 1M tokens. However, the metering pipeline emitted the 1K-scaled value (44) directly into the 1M meter:

  • date: 2026-10-19T00:00:00Z
  • meterName: Qwen38 27b SFT Gl 1M Tokens
  • meterId: 59a176f8-721a-5e7d-954a-250a4e3a5e3d
  • quantity: 44
  • unitOfMeasure: 1M
  • effectivePrice: 3
  • costInUSD: 132
  • costInBillingCurrency: 20796.6

Request to Microsoft Moderators

Because this is a clear platform metering defect (reporting 44 instead of 0.044 on a 1M token meter), a customer should not be forced to pay for a Technical Support plan just to get an erroneous $132.00 invoice corrected.

Could a Microsoft Moderator please:

  1. Enable a one-time free technical support ticket on my subscription, OR
  2. Escalate Service Request #2610200010000024 internally to the Azure AI Foundry / Billing Engineering team so this 1,000x overcharge ($131.87 USD / ¥20,775.80 JPY) can be credited?

(I have attached screenshots of the Microsoft Foundry job summary showing "Training tokens billed: 44,000" and the Cost Management analysis screen showing ¥20,797.)FOUNDRY-PORTAL

COST-MANAGEMENT

Cost Management
Cost Management

A Microsoft offering that enables tracking of cloud usage and expenditures for Azure and other cloud providers.

0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Jerald Felix 19,080 Reputation points Volunteer Moderator
    2026-10-29T17:18:58.9533333+00:00

    Hello Akihisa Hirosawa,

    Greetings! Thanks for raising this question in the Q&A forum

    Based on the records provided, this strongly indicates a metering unit conversion defect rather than valid consumption. The Foundry-reported 44,000 billable training tokens equal 0.044 units on a meter whose unit is one million tokens:

    44,000 / 1,000,000 × $3.00 = $0.132

    Recording a quantity of 44 against that meter produces $132.00, which is exactly 1,000 times higher. Microsoft’s fine-tuning documentation confirms that SFT training is token-based and that rates expressed per million tokens must be calculated using the token count divided by 1,000,000: Manage fine-tuning costs.

    1. Request reassignment of the existing billing case

    Reply to SR 2610200010000024 and ask that it be escalated as an incorrect metered quantity or usage-rating dispute, not as general Foundry technical troubleshooting. Request comparison of the originating Foundry usage event with the Azure commerce meter ingestion record.

    Include these identifiers:

    Meter: Qwen38 27b SFT Gl 1M Tokens

    Meter ID: 59a176f8-721a-5e7d-954a-250a4e3a5e3d

    SKU ID: DZH318Z0SDFC/01PV

    Region: eastus2

    Usage date: 2026-10-19

    Foundry billable tokens: 44,000

    Usage quantity received by billing: 44

    Expected meter quantity: 0.044

    Actual cost: $132.00

    Expected cost: $0.132

    Ask the agent not to close the case until the metering quantity has been validated by the Azure commerce or Foundry billing engineering team.

    Do not purchase a technical support plan for the billing dispute

    Azure provides unlimited subscription-management support, including billing, without requiring a paid technical support plan. This is stated in Microsoft’s official guidance: Create an Azure support request.

    If the current request is closed or incorrectly categorized, create another request through Azure portal > Help + support > Create a support request, select Billing, reference SR 2610200010000024, and attach the same evidence.

    Preserve the source evidence

    Attach the Foundry job summary, job ID, resource ID, UTC timestamps, full Microsoft.Consumption/usageDetails JSON, Retail Prices API response, Cost Analysis export, and invoice entry. These records allow Microsoft to determine whether the quantity was emitted at the wrong scale or transformed incorrectly during rating.

    Only Microsoft Billing Support can validate the backend usage event and approve a correction or credit. However, the arithmetic and meter metadata supplied here provide a well-supported basis for escalation and should not be dismissed merely because 44 × $3.00 matches the generated invoice.

    If this answer helps you kindly accept the answer which will help others who have similar questions.

    Best Regards,

    Jerald Felix.

    Was this answer helpful?

    2 people found this answer helpful.

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.