add azure ai grok 4 20 models

[emerzon]

Add support for Grok 4.20 models in Azure AI Foundry

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing emerzon:add-azure-ai-grok-4-20-models (2565f8e) with main (f74d626)}

[Copilot]

Pull request overview

Adds Azure AI Foundry pricing/capability metadata entries for the Grok 4.20 model variants so LiteLLM can resolve context limits, pricing, and feature flags for these model IDs.

Changes:

• Add azure_ai/grok-4-20-reasoning model entry (pricing, 256K context, reasoning/tooling flags, source link).
• Add azure_ai/grok-4-20-non-reasoning model entry (pricing, 256K context, tooling flags, source link).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR adds two new Azure AI Foundry model entries — azure_ai/grok-4-20-reasoning and azure_ai/grok-4-20-non-reasoning — to both the main and backup pricing/context-window JSON files, with 256K context windows and pricing matching the native xai variants ($2/$6 per M tokens). The backup file diff also includes unrelated OpenAI priority-tier pricing adjustments (gpt-5, gpt-5.4-pro, gpt-5.4-mini), which appear to be a file-sync rather than intentional new changes.

Two concerns raised in prior review rounds remain unaddressed: the supports_vision: true flag (present on these entries but absent from all other azure_ai/grok-* models) and the catalog-page source URLs rather than the official pricing page used by sibling entries.

Confidence Score: 4/5

Safe to merge if the author can confirm vision support and response-schema support on Azure AI Foundry for these specific model deployments.

Two P2 findings from prior review rounds — unverified supports_vision: true and catalog-page source URLs — remain open without author confirmation; the supports_response_schema concern is newly identified. All three are data-accuracy issues rather than runtime crashes, but incorrect capability flags can mislead callers into sending image content or expecting structured output on backends that won't honour them.

model_prices_and_context_window.json — capability flags (supports_vision, supports_response_schema) on the two new grok-4-20 entries need confirmation against the actual Azure AI Foundry deployment.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds azure_ai/grok-4-20-reasoning and azure_ai/grok-4-20-non-reasoning entries with 256K context windows, matching xai native pricing ($2/$6 per M tokens) but with supports_vision: true and catalog-page sources — both flags raised in prior review remain unaddressed.
litellm/model_prices_and_context_window_backup.json	Contains the same two new grok-4-20 entries plus additional out-of-scope changes: removes priority-tier cost fields from gpt-5/gpt-5.4-pro/gpt-5.4-pro-2026-03-05 and adds priority-tier fields to gpt-5.4-mini — these appear to be file-sync changes bringing the backup in line with the main file.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    xai["xai/grok-4.20-beta-0309-*\n(2M ctx, $2/$6 per M tokens\nsupports_vision: true\nNO supports_response_schema)"]
    azure4["azure_ai/grok-4\n(131K ctx, $3/$15 per M tokens\nno vision)"]
    azure41["azure_ai/grok-4-1-fast-*\n(131K ctx, $0.2/$0.5 per M tokens\nno vision)"]
    azure420r["azure_ai/grok-4-20-reasoning\n(256K ctx, $2/$6 per M tokens\nsupports_vision: true ⚠️\nsupports_reasoning: true\nsupports_response_schema: true)"]
    azure420nr["azure_ai/grok-4-20-non-reasoning\n(256K ctx, $2/$6 per M tokens\nsupports_vision: true ⚠️\nsupports_response_schema: true)"]

    xai -->|"Azure AI Foundry route"| azure420r
    xai -->|"Azure AI Foundry route"| azure420nr
    azure4 --- azure41
    azure41 --- azure420r
    azure420r --- azure420nr

Loading

_{Reviews (2): Last reviewed commit: "update backup model db" | Re-trigger Greptile}

[greptile-apps]

Source URL uses catalog page, not pricing documentation

The source for both new entries points to https://ai.azure.com/catalog/models/grok-4-20-*, while every other azure_ai/grok-* entry references an official pricing or blog page (e.g. https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/grok/). Catalog pages may change structure or become unavailable for models not yet publicly GA, making them a less stable authoritative source for pricing data. Consider linking to the Azure AI Foundry pricing page instead, consistent with azure_ai/grok-4 and azure_ai/grok-4-1-fast-*.

[greptile-apps]

First azure_ai grok entry to claim supports_vision: true

Every existing azure_ai/grok-* model (grok-3, grok-4, grok-4-1-fast, etc.) omits supports_vision or leaves it unset, while the native xai/grok-4.20-* entries set it to true. If Azure AI Foundry actually routes these models through an image-capable endpoint, the flag is correct. If not — for example, if the Foundry deployment strips vision inputs — liteLLM will silently forward image content to a backend that will reject it, confusing callers. Please confirm that grok-4-20-reasoning in Azure AI Foundry accepts multimodal inputs before merging.