feat(azure): add azure_ai/Microsoft-Decision-1

[berriai-litellm-provider-info-sync]

TLDR

Problem this solves:

• Azure lists Microsoft-Decision-1 in the Models API and bills it in the retail price feed, but the cost map has no row for it

How it solves it:

• Adds azure_ai/Microsoft-Decision-1 as a mode evaluation row priced from the Azure retail price feed
• Same edit in model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json

Rows

azure_ai/Microsoft-Decision-1 (new, kind page, source https://prices.azure.com/api/retail/prices?$filter=productName eq 'Microsoft Decision Models')

• input_cost_per_token 4.2e-08 and output_cost_per_token 4.2e-08: the feed has one meter for the model in all 37 regions, "productName":"Microsoft Decision Models","skuName":"Microsoft Decision 1","meterName":"Microsoft Decision 1 1M Tokens","retailPrice":0.042,"unitOfMeasure":"1M","armSkuName":"Pi 1 glbl", effective 2026-10-01. It has no input or output split, so the one token rate goes on both fields
• mode evaluation: the Azure Models API (Microsoft.CognitiveServices/locations/eastus2/models, Microsoft.Microsoft-Decision-1.1) states "capabilities": {"chatCompletion": "false", "decision": "true"}, and the other decision models in the catalog (typesafe/jev-latest, perplexity/pplx-decider-v1-27b, cloudflare/clef) use mode evaluation
• litellm_provider azure_ai: the Models API lists it under publisher Microsoft, format Microsoft, kind MaaS, the same as azure_ai/MAI-Thinking-1
• No max_input_tokens, max_output_tokens or supports_* flags: no Azure page or the Models API states a context window or capability flags for this model, and it is not on the Microsoft pricing page or models-sold-directly yet. Nothing is inherited, following the decision rows that also carry no limits (typesafe/jev-latest)
• No deprecation_date: the Models API gives "deprecation": {"inference": "2099-12-31T00:00:00Z"}, a placeholder rather than a retirement date

Changed pages

The Microsoft and Black Forest Labs pricing pages and models-sold-directly were re-read. The Microsoft and BFL prices render client side, so they were compared through the retail feed: MAI-Thinking-1 (2.00 / 0.20 cached / 8.00 per 1M) and MAI-Cyber-1-Flash (0.60 / 0.06 / 3.50) match the catalog, and models-sold-directly still states 256,000 context and 64,000 output for MAI-Thinking-1 and 32,000 tokens for the FLUX.2 and MAI-Image rows. azure_ai/flux.2-pro stays unchanged per ruling azure-flux-2-pro-megapixel-leave-unchanged

Delisted by the provider

Absent from the Models API and not dated on the retirement schedule or the retired models page, left for a human: azure/computer-use-preview, azure/container, azure/gpt-3.5-turbo, azure/gpt-3.5-turbo-instruct-0914, azure/gpt-35-turbo, azure/gpt-35-turbo-instruct, azure/gpt-35-turbo-instruct-0914, azure/gpt-4-0125-preview, azure/gpt-4-1106-preview, azure/gpt-4-turbo, azure/gpt-4-turbo-vision-preview, azure/gpt-5.2-pro, azure/gpt-5.5-pro, azure/gpt-5.6, azure/o1-mini, azure/sora-2-pro, azure/sora-2-pro-high-res. azure/mistral-large-2402 and azure/mistral-large-latest stay undated per ruling azure-mistral-large-conflicting-retirement-stay-undated and azure/ada keeps its date per azure-ada-keep-text-embedding-ada-002-date. The other delisted ids already carry a deprecation_date

Pre-Submission checklist

• I have added meaningful tests
• My PR's scope is as isolated as possible; it only solves 1 specific problem

---

Note

Low Risk
Pricing-catalog JSON only; no runtime routing, auth, or billing logic changes.

Overview
Adds azure_ai/Microsoft-Decision-1 to the LiteLLM model cost catalog so usage can be costed when Azure exposes this decision model via MaaS.

The new row uses mode: evaluation, litellm_provider: azure_ai, and the same input/output per-token rate (4.2e-08) from the Azure retail price feed. The change is mirrored in model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json.

^{Reviewed by Cursor Bugbot for commit 20b6b8d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Confidence Score: 5/5

[Medium impact] This catalog-only change appears safe to merge

Summary

Adds azure_ai/Microsoft-Decision-1 to both price catalogs with mode: evaluation and matching input and output rates of 4.2e-08 per token

• The price catalogs include Microsoft Decision 1 as an Azure AI evaluation model.

_{Reviews (1) · Last reviewed commit: "feat(azure): add azure_ai/Microsoft-Deci..." · Reviewed by Greptile}

[codspeed]

Merging this PR will not alter performance

⚠️ 1 benchmark spent significant time in system calls

System calls cannot be consistently instrumented, so they are not included in the measure, which understates the real cost. Please switch to the Walltime instrument to accurately measure system calls.

Measurement and system calls

• test_completion_streaming

✅ 31 untouched benchmarks

---

_{Comparing litellm-providers/price-sync-azure (20b6b8d) with main (fe955f5)¹}

Footnotes

1. No successful run was found on main (0519882) during the generation of this report, so fe955f5 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[kerrylu-berri]

LGTM

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[emerzon]

@kerrylu-berri - The entry is added but the whole implementation is missing. Also the price is incorrect, this model doesnt charge output tokens