A run of the group-travel team multiplies model calls (lesson 0.2), and several customers share one project, one Foundry resource and one monthly cap (R-09, M-09). Three questions remain: How many tokens may a run consume? Who owns what was consumed, so that a chargeback is possible (M-11)? And how available is the team when services fail? Amounts do not appear; where the pricing page is needed, it says so.
Prior knowledge. Tokens and their limits: "AI track, lesson T1.2". Monthly cap as code and Azure budgets as a warning: "T10.3". Quotas and measurement plan: "AI-103, lesson 1.3". Call counter, GrenzClient and the identifiers: "lesson 0.2"; gate and call limit: "1.4"; rate limits and times: "3.3"; spans: "3.4". Budgets and cost analysis: "K3, lesson 3.1.2". SLI, SLO, SLA and composite targets: "K3, lesson 3.15.1". None of this is repeated.
New in this lesson (assumptions of the task). (1) Besides 25 calls (1.4), a ceiling of 200,000 tokens (input plus output) applies per run. (2) A usage ledger has one entry per model call (mandantId, deployment, input tokens, output tokens). (3) Allocation: model usage by weights per deployment, shared costs by share of the runs, the non-allocable remainder shown. (4) Availability target: 99% of the requests are processed within 10 minutes (month); no second region.
The requirement
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 3.5-01 | A run ends at 25 calls or 200,000 tokens, whichever is reached first. | R | Limit in code | Must | R-09 |
| 3.5-02 | Every model call produces an entry with mandantId, deployment and tokens. |
F | 100% of the calls | Must | M-11 |
| 3.5-03 | The allocation of a month is reconciled with the total in Cost Management; what is not allocated appears as the remainder. | Q | Deviation explained | Must | M-11 |
| 3.5-04 | Reports name only the mandantId, no names. |
R | 0 names | Must | M-07 |
| 3.5-05 | 99% of the requests are processed within 10 minutes. | Q | Monthly value | Should | M-02 |
Cost frame. Cost drivers: tokens per million (input, output and reasoning tokens), reserved capacity if chosen, Application Insights per gigabyte, storage for exports. Prices: the pricing pages of Foundry and Azure Monitor.