fix(model_prices): add provider-announced deprecation dates for OpenAI, Gemini, Vertex, Azure models

[devin-ai-integration]

TLDR

Problem this solves:

• 83 registry entries lack provider-announced retirement dates
• text-embedding-004 carried a wrong retirement date
• Claude Fable/Mythos 5 missing structured-output flag

How it solves it:

• Adds verified deprecation_date from official retirement docs
• Corrects text-embedding-004 to Vertex's published date
• Flags supports_native_structured_output on 3 Anthropic models

User Flow

Before: an admin running gpt-4.1 on Azure OpenAI gets no warning that the deployment retires on 2027-04-14

1. They open https://litellm-domain/ui/?page=models and look at their azure/gpt-4.1 deployment
2. No retirement date is shown, because the registry has none for that model
3. They call GET https://litellm-domain/model/info and the returned model info for azure/gpt-4.1 has no deprecation date field
4. The retirement lands with no advance notice and calls start failing with provider "model not found" errors

After: the same admin sees the published retirement date for that deployment

1. They open https://litellm-domain/ui/?page=models and look at their azure/gpt-4.1 deployment
2. They call GET https://litellm-domain/model/info and the model info for azure/gpt-4.1 now carries "deprecation_date": "2027-04-14", matching Microsoft's published schedule
3. The same is true for their Azure AI Foundry (azure_ai/*), Vertex (gemini-2.5-*, veo-*, embeddings) and legacy OpenAI (babbage-002, davinci-002, gpt-3.5-turbo-instruct) deployments
4. They can plan migrations ahead of the retirement instead of discovering it from failed requests

Relevant issues

Supports #26900 (proactive model deprecation alerts and /model/deprecations endpoint), which consumes registry deprecation_date values.

Changes

All values below come from official provider retirement/lifecycle docs; every date was read from the raw HTML/Markdown of the source page, not a rendered summary. Provider-announced dates for direct APIs were not copied to cloud-hosted variants (or vice versa) — each entry is dated only from the doc that covers that platform.

OpenAI — https://platform.openai.com/docs/deprecations

Models	deprecation_date
gpt-3.5-turbo-instruct, babbage-002, davinci-002	2026-09-28
text-moderation-007, text-moderation-latest, text-moderation-stable	2025-10-27

Size/quality-derived keys (1024-x-1024/...) were left alone, matching existing registry convention.

Gemini API — https://ai.google.dev/gemini-api/docs/deprecations

Model	deprecation_date
gemini/gemini-robotics-er-1.6-preview	2026-08-31

The Gemini API page still shows "No shutdown date announced" for gemini-2.5-pro / -flash / -flash-lite, so no date was added to the gemini/* keys for those.

Vertex AI — https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions (Discontinuation date column)

Models	deprecation_date
gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite	2026-10-20
gemini-2.5-flash-image, vertex_ai/gemini-2.5-flash-image	2026-10-02
vertex_ai/veo-2.0-generate-001, vertex_ai/veo-3.0-generate-001, vertex_ai/veo-3.0-fast-generate-001	2026-06-30
text-embedding-005, text-multilingual-embedding-002, multimodalembedding@001	2027-04-01
text-embedding-004 (was 2026-01-14, not on the page)	2027-04-01

gemini-embedding-001 was left undated: the page says "No sooner than May 20, 2028", which is not a committed date.

Azure OpenAI + Azure AI Foundry — https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule (Retirement date column)

Azure OpenAI (azure/*): gpt-4.1 2027-04-14, gpt-4.1-mini 2027-04-14, gpt-4o-mini 2027-04-14, gpt-4.1-nano 2026-10-14, gpt-5/-mini/-nano 2027-02-09, gpt-5.1/-codex/-codex-mini 2027-05-15, gpt-5.2 2027-06-08, gpt-5.4 2027-09-02, gpt-5.4-mini/-nano 2027-09-21, gpt-5.4-pro 2027-09-07, gpt-5.5 2027-10-26, gpt-image-1.5 2027-06-16, gpt-image-2 2027-10-21, o1 2026-10-21, o3 2026-10-21, o3-mini 2026-10-01, o3-pro 2026-12-17, o4-mini 2026-10-16.

Azure AI Foundry (azure_ai/*): claude-opus-4-1 2026-08-05, claude-haiku-4-5/claude-opus-4-5/claude-sonnet-4-5 2026-10-19, claude-opus-4-6 2027-02-02, claude-sonnet-4-6 2027-02-10, claude-opus-4-7 2027-04-06, deepseek-r1 2026-08-13, deepseek-v3-0324/deepseek-v3.1 2026-07-13, deepseek-v4-flash/deepseek-v4-pro 2028-02-20, grok-3/grok-3-mini (incl. global/) and grok-4-fast-reasoning/-non-reasoning 2026-05-01, kimi-k2.5 2027-01-26, kimi-k2.6 2027-04-16, FW-DeepSeek-V3.2/FW-GLM-5/FW-GLM-5.1/FW-Kimi-K2.5/FW-MiniMax-M2.5 2027-07-01, Llama-3.2-11B-Vision-Instruct/Llama-3.2-90B-Vision-Instruct/Meta-Llama-3.1-405B-Instruct/Meta-Llama-3.1-8B-Instruct 2026-06-13, cohere-rerank-v3.5 2026-05-14, mistral-document-ai-2505 2026-07-20, MAI-Image-2e 2026-08-15, gpt-5.4/gpt-5.4-2026-03-05 2027-09-02, gpt-5.4-mini(+snapshot)/gpt-5.4-nano(+snapshot) 2027-09-21, gpt-5.4-pro(+snapshot) 2027-09-07, gpt-5.5 2027-10-26.

Models with several live versions under one alias (e.g. azure/gpt-4o, azure/sora-2, azure_ai/model_router) were skipped rather than pinned to one version's date.

Anthropic capability flag — https://docs.claude.com/en/docs/build-with-claude/structured-outputs

claude-fable-5, claude-mythos-5, claude-mythos-preview are listed as supported models for structured outputs but were missing supports_native_structured_output. Added on the Anthropic-direct entries only (matching #35930, which did the same for claude-sonnet-5 / claude-haiku-4-5).

Audited, deliberately unchanged

• Anthropic pricing/context/cache minimums, Bedrock lifecycle dates, xAI, Groq, Cohere, Mistral, Fireworks: registry already matches the official docs.
• eu.anthropic.claude-opus-4-1-20250805-v1:0: the Bedrock lifecycle row lists only us-east-1/us-east-2/us-west-2, so the EU inference-profile key was left undated.
• DeepSeek deepseek-v4-flash / deepseek-v4-pro pricing: https://api-docs.deepseek.com/quick_start/pricing now lists only peak ($0.44/$1.32 in, $1.32/$3.96 out per 1M) and off-peak (half those) rates, while the registry pins $0.14/$0.28 and $0.435/$0.87 as a "75% discounted active price" (test_deepseek_v4_models_in_cost_map, from [Feature]: Add cost mapping for Deepseek V4 Flash and Pro #26709). The registry matches neither published rate, but correcting it means changing that pinning test and choosing a single rate for a time-of-day price, so it is left for maintainers to decide.

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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

Existing coverage applies (this is a data-only change): tests/test_litellm/test_model_prices_schema.py (19 passed), tests/test_litellm/test_utils.py (264 passed), and python ci_cd/check_files_match.py confirming the root registry and litellm/model_prices_and_context_window_backup.json stay in sync.

Proof

Data-only change; no runtime behavior to exercise e2e. Verification of the change against the provider docs, at commit 0897b6d:

1. python3 -c "import json; d=json.load(open('model_prices_and_context_window.json')); print(d['azure/gpt-4.1']['deprecation_date'], d['gemini-2.5-pro']['deprecation_date'], d['babbage-002']['deprecation_date'])" → 2027-04-14 2026-10-20 2026-09-28, matching the Microsoft, Vertex and OpenAI pages linked above.
2. python3 ci_cd/check_files_match.py → Passed! Files model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json match.
3. LITELLM_LOCAL_MODEL_COST_MAP=True pytest tests/test_litellm/test_model_prices_schema.py tests/test_litellm/test_utils.py -q → 283 passed.

Link to Devin session: https://app.devin.ai/sessions/9b4d808bfacd40a1abb0c894c05e04a0

---

Note

^{Cursor Bugbot is generating a summary for commit 0897b6d. Configure here.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[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]

Greptile Summary

The PR synchronizes provider-announced model retirement dates into both model registries and enables native structured-output metadata for three Anthropic models.

• Adds retirement metadata for selected OpenAI, Gemini, Vertex AI, Azure OpenAI, and Azure AI Foundry models.
• Corrects the retirement date for text-embedding-004.
• Keeps the primary and backup model registries synchronized.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds and corrects model lifecycle metadata and three structured-output capability flags; no eligible follow-up defect was established.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes without introducing divergence.

_{Reviews (2): Last reviewed commit: "fix(model_prices): add provider-announce..." | Re-trigger Greptile}

[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 litellm_model_registry_lifecycle_audit_20260817 (0897b6d) with litellm_internal_staging (2bc87ec)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (b69068c) during the generation of this report, so 2bc87ec was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 0897b6d. Configure here.}

[devin-ai-integration]

Superseded by #37283, which adds the same provider-announced deprecation dates (identical values on the 51 overlapping entries) plus 122 more. Closing in favor of that PR.