fix(cost): stop double-billing the Azure AI Model Router flat fee

[Anuj7411]

TLDR

Problem this solves:

• Azure AI Foundry Model Router requests are billed the flat routing fee twice
• logged spend for those requests is ~2x the real cost

How it solves it:

• when the token-cost path already folded the flat fee into prompt cost, skip adding it again via additional_costs

User Flow

Before: a developer routing through Azure AI Foundry Model Router sees double the expected spend

1. They send a completion whose model resolves to the Azure Model Router (e.g. azure-model-router)
2. cost_per_token folds the flat $0.14 per million input-token fee into prompt cost, and completion_cost then adds the same fee again through its additional_costs hook
3. Their logged spend is twice the flat fee

After: the same request is billed the flat fee once

1. They send the same completion
2. The flat fee is billed a single time through the token-cost path
3. Logged spend matches the real charge

Relevant issues

Type

🐛 Bug Fix

Changes

completion_cost computes token cost via azure_ai cost_per_token, which already folds the Model Router flat fee into prompt cost whenever it detects the router from the response model or the request_model (is_router_request). Right after, for custom_llm_provider == "azure_ai", completion_cost separately computed the same flat fee through _get_additional_costs and summed it into the total, so any request where the token path detected the router was charged the fee twice. This guards the additional_costs branch: when _is_azure_model_router(model) or _is_azure_model_router(request_model_for_cost) is true, the token path already billed the fee, so additional_costs is left unset. The hidden-params-only path (response carries the underlying model, router known only from _hidden_params) is unchanged and still bills the fee once through additional_costs.

Screenshots / Proof of Fix

For a pure azure-model-router response with 5000 prompt tokens, completion_cost now returns a single flat fee (5000 * rate / 1e6) instead of twice that. Regression test tests/test_litellm/llms/azure_ai/test_azure_ai_cost_calculator.py::TestAzureModelRouterCostBreakdown::test_additional_costs_in_cost_breakdown asserts the single-fee total; it fails on the pre-fix code (double) and passes after. All 32 tests in that file pass.

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Anuj7411:litellm_fix_azure_model_router_double_billed_flat_fee (bbe7aea) with litellm_internal_staging (b091158)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (d2d158f) during the generation of this report, so b091158 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

Closing as superseded: #40189 landed the same model-router flat-fee guard in completion_cost, proven live with the fee charged once. Thanks for the contribution!