fix(model_prices): annotate retired xAI grok-2-era models and mark grok-4.20-multi-agent Responses-only

[ben7am1n]

TLDR

Problem this solves:

• xAI retired the grok-2 generation (effective 2026-02-28) and the retired slugs now hard-fail with "Model not found — retired by xAI", but LiteLLM's model list gave no hint
• xai/grok-4.20-multi-agent-* was listed as a chat model even though xAI documents that it does not work with the Chat Completions API

How it solves it:

• Annotate the retired grok-2-era entries with deprecation_date so they surface via GET /v1/model/deprecations; pricing metadata is kept for historical spend data
• Mark the two grok-4.20-multi-agent entries as Responses-API-only (mode: responses)

User Flow

Before: a stale model entry causes unexplained production failures

1. A developer configures xai/grok-2 in their LiteLLM setup
2. Every request fails upstream with "Model not found — retired by xAI", and GET /v1/model/deprecations lists nothing about it
3. The admin cannot tell which of their configured models are dead upstream

After: retirement is visible before it breaks anything

1. The same developer configures the same model on the updated gateway
2. GET /v1/model/deprecations lists xai/grok-2 (and the other grok-2-era slugs) with its deprecation date
3. Browsing models, xai/grok-4.20-multi-agent-0309 shows under Responses-API capabilities rather than chat

Relevant issues

Fixes #38179

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Caveats

• The May 15, 2026 retirement slugs were already annotated upstream; this PR covers the grok-2-era batch, whose official announcement (Feb 3, 2026 notice, effective 2026-02-28) matches the existing xai/grok-2-vision-1212 annotation
• xai/grok-beta / xai/grok-vision-beta were retired earlier without a published date; they are confirmed hard-failing upstream (repro in [Bug]: xAI models can't use #38179) and are annotated with the same batch date — happy to adjust per maintainer preference
• Provider-hosted variants (oci/xai.grok-4.20-multi-agent etc.) are untouched: the Responses-only limitation is xAI-API-specific

Screenshots / Proof of Fix

tests/test_litellm/test_xai_grok_model_metadata.py — 4 passed
- retired grok-2-era entries carry deprecation_date 2026-02-28
- grok-4.20-multi-agent entries are mode=responses
- active models (grok-4.5/4.6/4.20-0309) are NOT marked deprecated
- main and backup cost maps stay in sync

[greptile-apps]

Greptile Summary

This PR annotates retired xAI Grok model entries with deprecation dates and marks two Grok multi-agent models as Responses-only

• Updates the primary and backup model metadata maps consistently
• Adds regression coverage for retirement dates, API modes, active models, and map parity
• The metadata changes align with the existing deprecation-reporting and Responses dispatch contracts

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking request to place the regression coverage in the mapped tests and remove unused global mutation

The model metadata updates are synchronized and match supported deprecation and Responses-mode behavior; the only accepted concern is test organization and unnecessary fixture state mutation

Files Needing Attention: tests/test_litellm/test_xai_grok_model_metadata.py

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds valid ISO deprecation dates and changes two xAI multi-agent entries to the supported responses mode without an identified functional defect
litellm/model_prices_and_context_window_backup.json	Mirrors the primary model metadata changes and remains synchronized for all affected entries
tests/test_litellm/test_xai_grok_model_metadata.py	Adds useful metadata assertions, but creates a separate bug-fix test module and includes an unnecessary global model-cost fixture

_{Reviews (1): Last reviewed commit: "fix(model_prices): annotate retired xAI ..." | Re-trigger Greptile}

[greptile-apps]

Avoid redundant test state

This bug fix creates a separate test module whose autouse fixture replaces litellm.model_cost, although every assertion reads the JSON files directly. This bypasses the mapped-test convention and adds global-state and cache-isolation risk without contributing to the regression coverage

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing ben7am1n:fix/xai-retired-models (f0d753b) with litellm_internal_staging (8ebcb3e)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (cd63c7e) during the generation of this report, so 8ebcb3e was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[devin-ai-integration]

Folded into rolling PR #38693; the Feb 28 dates could not be verified on live xAI docs, other changes already landed.