fix(model_prices): correct stale retirement dates (Azure, Vertex, Bedrock) and add deepseek-v4-flash-vision-exp

[devin-ai-integration]

TLDR

Problem this solves:

• Four Azure model retirement dates were stale or wrong
• azure/gpt-4o, azure/sora-2, TTS/transcribe entries had no retirement date
• text-embedding-004 retirement date predated Google's current schedule
• eu.anthropic.claude-opus-4-1 was missing its Bedrock EOL date
• DeepSeek's vision model was missing from the registry

How it solves it:

• Set each deprecation_date to the provider's published retirement date
• Added deepseek-v4-flash-vision-exp (and deepseek/ alias)
• Every value taken from the official provider table (links below)

User Flow

Before: a proxy admin tracking model end-of-life sees wrong dates for Azure GPT-4.1-nano and Vertex text-embedding-004, and no date at all for azure/gpt-4o

1. They call GET https://litellm-domain/model/deprecations
2. azure/gpt-4.1-nano comes back with "deprecation_date": "2026-10-14" — six months earlier than Microsoft's published retirement, so it looks urgent when it isn't
3. text-embedding-004 comes back with "deprecation_date": "2026-01-14", a date already in the past, so the model reads as retired while it still serves traffic
4. azure/gpt-4o, azure/sora-2, azure/gpt-4o-mini-tts, azure/gpt-4o-mini-transcribe and eu.anthropic.claude-opus-4-1-20250805-v1:0 are absent from the response entirely, so no advance warning fires for them
5. They send POST https://litellm-domain/v1/chat/completions with "model": "deepseek/deepseek-v4-flash-vision-exp" and the response is logged at $0 spend, since the model has no pricing entry

After: the same endpoint reports each provider's published retirement date, and the DeepSeek vision model is priced

1. They call GET https://litellm-domain/model/deprecations
2. azure/gpt-4.1-nano now reports 2027-04-14, matching Microsoft's schedule
3. text-embedding-004 now reports 2027-04-01, matching Google's schedule
4. azure/gpt-4o (2027-04-14), azure/sora-2 (2026-09-15), azure/gpt-4o-mini-tts and azure/gpt-4o-mini-transcribe (2027-06-15) and eu.anthropic.claude-opus-4-1-20250805-v1:0 (2027-01-08) now appear, so alerts fire ahead of retirement
5. The same POST with "model": "deepseek/deepseek-v4-flash-vision-exp" is logged at real spend at $0.44/M input, $1.32/M output, and image input is advertised as supported

Sources for every changed value

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

key	before	after	doc row
azure/gpt-4.1-nano	2026-10-14	2027-04-14	gpt-4.1-nano 2025-04-14, Legacy, 2027-04-14
azure/gpt-4.1-nano-2025-04-14	2026-10-14	2027-04-14	same row
azure/us/gpt-4.1-nano-2025-04-14	2026-10-14	2027-04-14	same row
azure/gpt-4o	(none)	2027-04-14	gpt-4o 2024-08-06 and 2024-11-20, 2027-04-14
azure/gpt-4o-mini-transcribe	(none)	2027-06-15	gpt-4o-mini-transcribe 2025-12-15, GA, 2027-06-15
azure/gpt-4o-mini-tts	(none)	2027-06-15	gpt-4o-mini-tts 2025-12-15, GA, 2027-06-15
azure/sora-2	(none)	2026-09-15	sora-2 2025-12-08, Preview, 2026-09-15

The earlier 2026-10-14 value matches no row on that page; the fine-tuned-model table on the lifecycle page (2027-04-01/2027-04-14 training, 2027-10-xx deployment) covers fine-tuned deployments only and is not the base-model retirement date.

Vertex AI — https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions ("Embeddings models" table): text-embedding-004, released May 14 2024, retirement April 1 2027 → text-embedding-004 2026-01-14 → 2027-04-01. Sibling rows text-embedding-005, text-multilingual-embedding-002, multimodalembedding@001 already carry 2027-04-01 in the registry.

Bedrock — https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html (legacy/EOL table): anthropic.claude-opus-4-1-20250805-v1:0 EOL January 8 2027 → added to eu.anthropic.claude-opus-4-1-20250805-v1:0, which lacked it while the base and us. keys already had it.

DeepSeek — https://api-docs.deepseek.com/quick_start/pricing: deepseek-v4-flash-vision-exp shares the deepseek-v4-flash column values (1M context, 384K max output, peak $0.014/M cache-hit input, $0.44/M cache-miss input, $1.32/M output; JSON output, tool calls, chat-prefix completion supported). New entries deepseek-v4-flash-vision-exp and deepseek/deepseek-v4-flash-vision-exp mirror the existing flash entries with supports_vision: true.

Audited and left unchanged

Checked against the raw provider tables, no registry change needed:

• Anthropic https://docs.claude.com/en/docs/about-claude/model-deprecations — all 16 dated rows already match, including Fable 5, Opus 5/4.8/4.7/4.6, Sonnet 5/4.6. claude-mythos-preview is stated as deprecated with no retirement date published, so nothing to write.
• Gemini API https://ai.google.dev/gemini-api/docs/deprecations and Vertex model-versions — every announced shutdown date already matches (gemini-3.1-flash-lite 2027-05-07, gemini-2.5-flash-image 2026-10-02, gemini-embedding-001 2028-05-14 on the Gemini key vs 2028-05-20 on the Vertex key, imagen-4.0 2026-08-17, veo 2026-06-30, robotics-er-1.6 2026-08-31). The GA Gemini 3.x models say "no shutdown date announced" and are correctly left without one.
• OpenAI https://platform.openai.com/docs/deprecations — all shutdown rows match. Not changed: gpt-4o-realtime-preview*, gpt-4o-mini-realtime-preview, gpt-4o-audio-preview, gpt-4o-mini-audio-preview carry 2027-01-20 (the gpt-4o-realtime/gpt-4o-audio family row) while the older per-snapshot rows say 2026-05-07; the page supports both readings, so left alone rather than guessed.
• Azure partner tables (Anthropic, DeepSeek, MoonshotAI, Fireworks, Microsoft) — all already match. azure/gpt-realtime-mini-2025-10-06 has two conflicting rows for the same version (2027-04-06 and 2026-09-21) so it was left unset.
• Bedrock lifecycle table — all other rows already match.
• Cohere https://docs.cohere.com/docs/models — the Sept 15 2025 deprecations and existing context windows all match. command-a-plus-05-2026, command-a-reasoning-08-2025, command-a-vision-07-2025, command-a-translate-08-2025, rerank-v4.0-pro/fast are missing under the cohere provider, but Cohere's pricing page renders its per-model API rates client-side and no official price for them could be read, so they were not added. Same for xAI's Imagine image/video and voice models and Cohere's Aya/transcribe models.

Rolling registry PR

This is the single open registry PR. The four other open PRs touching this file are code/behavior changes that only incidentally edit the JSON, so they were left alone and nothing was closed: #37375 and #35987 (Bedrock Mantle web-search transformation fix) and #36422 and #36416 (moving fallback_generalizations out of the cost map).

Relevant issues

Supports #26900 (proactive model deprecation alerts / /model/deprecations).

Linear ticket

Pre-Submission checklist

• I have added meaningful tests — data-only change; covered by the existing schema/deprecation suites
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible
• I have received a Greptile Confidence Score of at least 4/5

Screenshots / Proof of Fix

After (36837d3)

Registry checks

1. 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.
2. uv run --no-sync python ci_cd/generate_model_prices_schema.py → schema regenerated with no diff
3. LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync pytest tests/test_litellm/test_model_prices_schema.py tests/test_litellm/test_model_cost_aliases.py tests/test_litellm/proxy/test_model_deprecations_endpoint.py tests/test_litellm/proxy/common_utils/test_model_deprecation.py tests/test_litellm/integrations/SlackAlerting/test_model_deprecation_alert.py -q → 66 passed

Link to Devin session: https://app.devin.ai/sessions/cee732482ae349b08b2b8c336a38e369

[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

This PR updates stale or missing retirement metadata for selected Azure, Vertex AI, and Bedrock models and registers DeepSeek V4 Flash Vision Experimental under bare and provider-prefixed names.

• Corrects three Azure GPT-4.1 Nano retirement dates.
• Adds retirement dates for Azure GPT-4o, audio models, Sora 2, and an EU Bedrock Claude model.
• Corrects the Vertex AI text-embedding-004 retirement date.
• Adds DeepSeek vision-model pricing, limits, capabilities, and alias metadata.
• Keeps the primary and packaged backup registries synchronized.

Confidence Score: 5/5

The PR appears safe to merge; no concrete changed-code failure was identified in the synchronized metadata updates.

The retirement dates use valid registry fields, the new DeepSeek entries follow established bare and provider-prefixed lookup conventions, and both runtime and source registry copies contain matching changes.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates retirement metadata and adds internally consistent DeepSeek vision-model registry entries without an identified defect.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes so packaged runtime metadata remains synchronized.

_{Reviews (1): Last reviewed commit: "fix(model_prices): correct stale retirem..." | 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 devin/1787318013-registry-audit (36837d3) with litellm_internal_staging (ff02d5c)}

[jasondelio]

Hi @mateo-berri , I just want to let you know this PR not complete for supporting deepseek-v4-flash-vision-exp.

Adding "supports_vision": True alone does not fix image inputs at runtime because DeepSeekChatConfig._transform_messages in litellm/llms/deepseek/chat/transformation.py unconditionally flattens content lists to strings before dispatching:

# litellm/llms/deepseek/chat/transformation.py
def _transform_messages(
    self, messages: list[AllMessageValues], model: str, is_async: bool = False
) -> list[AllMessageValues] | Coroutine[Any, Any, list[AllMessageValues]]:
    # ⚠️ This line unconditionally converts content lists to plain string,
    # stripping all `image_url` / `input_image` blocks!
    messages = handle_messages_with_content_list_to_str_conversion(messages)

Inside handle_messages_with_content_list_to_str_conversion, convert_content_list_to_str only extracts c.get("text") and silently drops any image_url objects. As a result, the outgoing request sent to DeepSeek only receives text and no images.

[devin-ai-integration]

Superseded by rolling registry audit #37902, which reverified and absorbed the provider-documented metadata changes at field level