AI Model Deprecation Calendar

Legacy across 31 regions from 2026-10-08, extended-access pricing from 2027-01-08, end of life 2027-04-08. The Claude API retires the same model on 2026-11-30, so Bedrock keeps it more than four months longer.

Part of a large 2026-09-30 Legacy batch, all with end of life on 2027-03-30 and no extended-access date announced yet. Azure Foundry retired its DeepSeek-R1 copy on 2026-08-13.

Exposed as deepseek.v3-v1:0 on bedrock-runtime and deepseek.v3.1 on bedrock-mantle, so a migration has to cover both ids. Part of the 2026-09-30 Legacy batch.

Part of the 2026-09-30 Legacy batch. Bedrock put the whole Llama 3 and Llama 4 line on notice at once, which leaves no Meta model in Bedrock's Active list on this page.

Part of the 2026-09-30 Legacy batch, same dates as Maverick.

Part of the 2026-09-30 Legacy batch, along with Llama 3 8B and 70B and Llama 3.1 8B and 70B. Going Legacy also blocks new fine-tuning jobs and new Provisioned Throughput on it.

Listed as moonshot.kimi-k2-thinking on bedrock-runtime and moonshotai.kimi-k2-thinking on bedrock-mantle. Part of the 2026-09-30 Legacy batch.

zai.glm-4.7-flash reaches end of life the same day. Part of the 2026-09-30 Legacy batch.

The 30B A3B sibling is on the same notice. Part of the 2026-09-30 Legacy batch.

Part of the 2026-09-30 Legacy batch, which also covers Mixtral 8x7B Instruct, Devstral 2 123B, Voxtral Mini 3B 2507 and Voxtral Small 24B 2507. That is every Mistral model on this page, so Bedrock has no Mistral model left in its Active list here.

Part of the 2026-09-30 Legacy batch, along with Gemma 3 12B IT and Gemma 3 4B IT. AWS lists the 27B entry under the model name Gemma 3 27B PT while the id is the instruction-tuned google.gemma-3-27b-it.

MiniMax M2 is on the same notice with the same dates, so both MiniMax models go together. Part of the 2026-09-30 Legacy batch.

The vision sibling, Nemotron Nano 12B v2 VL BF16 (nvidia.nemotron-nano-12b-v2), carries the same dates. Part of the 2026-09-30 Legacy batch.

Video understanding model, Legacy from 2026-09-22 across 30 regions with no extended-access date listed. Marengo Embed 3.0 shares the same dates in ap-northeast-2, which puts both the current TwelveLabs models on notice while the older Marengo Embed 2.7 ends first on 2026-11-30.

Legacy from 2026-09-21 in us-east-1, us-east-2, us-west-1 and us-west-2, with no extended-access date listed. Palmyra X4 went Legacy twelve days earlier with end of life on 2027-03-09, so both Writer models on Bedrock are now on notice.

In public extended access since 2026-10-08, where Anthropic can set higher prices. Already retired on the Claude API (2026-08-05).

Video embedding model. Extended-access pricing from 2026-08-29.

Extended-access pricing from 2026-08-26.

Extended-access pricing from 2026-08-26.

Regional batch: Legacy from 2026-10-10 in ap-south-1, ap-southeast-1 and eu-west-2, from 2026-10-13 in sa-east-1 and from 2026-10-15 in ca-central-1, all with end of life on 2026-11-13. Extended access starts on each region's Legacy date, so the usual three-month buffer does not apply here. The Claude API retired this model on 2025-07-21.

Legacy from 2026-10-10 in ap-northeast-2, ap-south-1 and ap-southeast-1 with end of life on 2026-11-13. Extended access starts the same day as Legacy. The Claude API retired it on 2025-10-28.

Same regions and dates as the June 2024 snapshot.

Legacy from 2026-10-10 in ap-south-1 and eu-west-2 with end of life on 2026-11-13. The Claude API retired it on 2026-02-19, so Bedrock carried it nine months longer.

Now in the Bedrock reached-end-of-life table, four months after the Claude API retired it on 2026-06-15. Public extended access ran from 2026-07-14.

Now in the Bedrock reached-end-of-life table. The lifecycle table lists no extended-access start date.

amazon.nova-reel-v1:1 reached end of life the same day. Both are now in the Bedrock reached-end-of-life table, which leaves Bedrock with no first-party Amazon image or video model.

End of life on 2026-09-14. The lifecycle table lists no extended-access start date.

End of life on 2026-09-14. The lifecycle table lists no extended-access start date.

End of life in the main commercial regions and GovCloud. Extended-access pricing applied from 2026-06-10. A second regional batch (ap-northeast-2, ap-south-1, ap-southeast-1, eu-central-2, sa-east-1, ca-central-1) went Legacy in October 2026 with end of life on 2026-11-13, so the model id still works in those regions for now.

The largest Llama Bedrock hosted, Legacy from 2026-01-07 in us-east-2 and us-west-2 with extended-access pricing from 2026-04-07. The whole Llama 3.2 line (1B, 3B, 11B and 90B Instruct) reached end of life the same day.

Anthropic notified developers on September 30, 2026 and retires the model on November 30, 2026. The dates apply to Anthropic-operated platforms (Claude API, Claude Platform on AWS, Microsoft Foundry); Amazon Bedrock and Google Cloud set their own schedules.

Tentative retirement date on an Active model, listed as not sooner than November 24, 2026. Microsoft Foundry now lists the same 2026-11-24 date for its hosted copy.

Tentative date that has now passed: Anthropic still lists the model as Active with a retirement not sooner than October 15, 2026, so treat it as live until a dated notice lands. Microsoft Foundry lists 2026-11-15 for both of its hosted versions. This is the model Anthropic names as the replacement on every retired Haiku row above.

Retired on the Claude API and Microsoft Foundry on 2026-08-05 (Foundry names claude-opus-5 as its replacement). Bedrock keeps it in extended access until 2027-01-08.

Claude API date. Bedrock put it in extended access from 2026-07-14 and ended support on 2026-10-14.

Invitation-only research preview. Anthropic lists it as deprecated but has not published a retirement date; Microsoft Foundry shows 2027-04-02 for its hosted copy. claude-fable-5 is the same capability without access gating.

Claude API date. The Google Cloud copy ran until 2026-08-23 and the Bedrock copy until 2026-09-10.

Original cost-tier Claude. Retired with the rest of the Claude 1 family (claude-1.0 through claude-1.3 and the earlier Instant versions).

Cohere words this notice as retired effective April 4, 2026, not deprecated, so there was no existing-customer grace period. embed-english-light-v2.0 is on the same notice; Cohere also names embed-english-v3.0. Changing embedding model means re-embedding the corpus, so treat this as a project, not a config change.

Retired effective April 4, 2026. Cohere also names embed-multilingual-v3.0, and Embed 5 Pro and Fast arrived on 2026-09-30.

Cohere Labs multilingual model, retired effective April 4, 2026. Cohere also names command-r7b-12-2024 and command-a-reasoning-08-2025.

Retired alongside c4ai-aya-expanse-8b, effective April 4, 2026. The named replacements are text models, so vision workloads need a different target; Embed 5 handles images on the retrieval side.

Deprecated in the September 2025 batch with command-light, the legacy generate and summarize endpoints, and fine-tuning. Cohere has not published a shutdown date; deprecated models stay open to existing users until one is assigned.

No shutdown date published.

The command-r alias and command-r-03-2024 are deprecated. Cohere suggests command-r-08-2024 or command-a-03-2025. Bedrock retired its Command R copy on 2026-08-19.

The command-r-plus alias and command-r-plus-04-2024 are deprecated. Cohere suggests command-r-plus-08-2024 or command-a-03-2025. Bedrock retired its Command R+ copy on 2026-08-19.

V4 Flash was retired when V4.1 Flash launched. The deepseek-v4-flash name is still accepted but served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price, so calls succeed and hit a different model. deepseek-flash is the supported name for the current Flash model. V4 Pro is unaffected.

Experimental vision model launched 2026-08-21 and retired three weeks later. The name is still accepted and routes to DeepSeek-V4.1-Flash-1017, which has native vision, billed at the Flash price.

Legacy alias stopped serving on 2026-07-24 at 15:59 UTC. Same base URL and key, so migration is a one-line model parameter change.

Legacy alias stopped serving on 2026-07-24 at 15:59 UTC. Mapping trap: deepseek-reasoner routed to V4 Flash in thinking mode, not to the stronger V4 Pro, so a naive swap to deepseek-v4-pro changes cost. Teams running heavy reasoning on the old alias should pick the tier deliberately.

Earliest possible Gemini API shutdown. Vertex AI lists no sooner than May 20, 2028. Switching embedding models means re-embedding the corpus, so the long runway matters.

Earliest possible shutdown on the Gemini API; Vertex AI lists May 7, 2027 or later. A lifecycle date rather than an active notice.

The date moved out by more than five months, from 2026-10-02 to 2027-03-15, so anyone who migrated on the old date gained nothing but anyone still on it has room. The Gemini API table now names Gemini 3.1 Flash-Lite Image directly.

Google names gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts. One of five speech and live models sharing a 2026-11-17 date.

Retires with gemini-2.5-pro-preview-tts on the same date. Google names gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts.

No Pro-tier speech successor is named; both replacements are Flash-class.

Live API model. gemini-3.8-live and gemini-3.8-live-extended-thinking are the current pair, both released 2026-09-15.

The last native-audio preview on the 2.5 line. Native audio dialogue moves to the Live API models.

Managed agent rather than a model, listed in the Managed Agents section of the Gemini API deprecations page alongside the Antigravity previews. Past its date but still listed; deep-research-max-preview-04-2026 is the other surviving research agent.

Video generation preview released 2026-06-30 and replaced by the GA model two months later. The shutdown date moved out from 2026-09-30 to 2026-10-22, the same day the three Veo 3.1 preview ids go.

Past its scheduled 2026-10-22 shutdown but still carrying that date on the deprecations table, so plan as if it is gone. Google retired the Veo preview ids on the Gemini API and routes video generation to Gemini Omni instead. GA Veo 3.1 ids remain on Google Cloud with a retirement no sooner than 2026-11-17, so a Gemini API caller and a Vertex AI caller face different migrations.

Same date as the standard and lite Veo 3.1 previews, now passed but still listed.

Released 2026-03-31, so under seven months of life. The cheapest Veo tier on the Gemini API at $0.05 per second of 720p. Past its date but still listed.

No shutdown date. API requests naming gemini-3.7-flash are routed automatically to gemini-3.8-flash, and the model is no longer selectable in Google AI Studio.

No shutdown date. API requests naming gemini-3.5-flash are routed automatically to gemini-3.6-flash; Google also lists gemini-3.8-flash as a migration option.

Managed agent rather than a model, listed in the Managed Agents section of the Gemini API deprecations page. The September 2026 preview that replaces it shipped on 2026-09-17, leaving about three weeks of overlap.

Past the scheduled date but not yet marked shut down.

Google is folding image generation into the Gemini image models, and the table now names gemini-nano-banana-2.1 rather than gemini-3.1-flash-image. Past the scheduled date but not yet marked shut down on the Gemini API; Vertex AI retired its copy on 2026-06-30.

Past the scheduled date but not yet marked shut down on the Gemini API; Vertex AI retired its copy on 2026-06-30. The table now names gemini-nano-banana-2.1.

Past the scheduled date but not yet marked shut down. The table now names gemini-nano-banana-2.1.

Multimodal embedding preview released 2026-03-10. The deprecations table lists it as embedding-2-preview. Vectors from a preview model should be re-embedded, not mixed with GA output.

The only computer-use model Google shipped on the Gemini API, released 2025-10-07 and shut down nine months later. The named replacement is a general Flash model, not a dedicated computer-use one.

Fifteen days of notice. The Gemini API table names gemini-omni-1.1-flash or the GA models on the Gemini Enterprise Agent Platform; the Vertex AI retired table names veo-3.1-generate-001.

Past its 2026-06-30 date on the Gemini API and already in the Vertex AI retired table. The Gemini API now names gemini-omni-1.1-flash; Vertex AI names veo-3.1-generate-001 as the GA replacement.

Past its 2026-06-30 date on the Gemini API and already in the Vertex AI retired table. The Gemini API now names gemini-omni-1.1-flash; Vertex AI names veo-3.1-fast-generate-001 as the GA replacement.

Past its scheduled shutdown but not yet marked shut down on the deprecations table. Plan as if it is gone. The table now points at gemini-nano-banana-2.1 rather than the GA gemini-3.1-flash-image.

Past its scheduled shutdown but not yet marked shut down on the deprecations table. Plan as if it is gone.

Shut down on the Gemini API and retired on Vertex AI the same day, including gemini-2.0-flash-001. The Gemini API table names gemini-3.6-flash; the Vertex AI retired table now names gemini-3.5-flash-lite.

Includes gemini-2.0-flash-lite-001. Retired on Vertex AI the same day.

First-generation Gemini API. Replaced by 1.5 family with longer context.

Original PaLM 2 text generation API. Migration to Gemini API required.

Retires alongside text-embedding-004 and multimodalembedding@001 on Vertex AI. No replacement named in the table; gemini-embedding-2 is the newest Google embedding model listed.

No replacement named in the table.

Vertex AI gives this model a dated retirement where the Gemini API gives none and instead routes requests to gemini-3.8-flash automatically, so a Vertex deployment needs a real migration and a Gemini API caller does not.

Native-audio Live API model on Google Cloud. The Gemini API equivalent, gemini-2.5-flash-native-audio-preview-12-2025, goes earlier on 2026-11-17.

Released 2026-07-21 and retiring on Vertex AI less than four months later. The Gemini API lists no shutdown date for gemini-3.6-flash and still names it as the replacement for Gemini 2.0 Flash and 3.5 Flash, so the two platforms disagree about whether this model has a future.

Listed as November 17, 2026 or later, so a lifecycle floor rather than a firm shutdown. veo-3.1-fast-generate-001 carries the same date. These are the GA ids the Gemini API points Veo users at, and they outlive the Gemini API preview ids that went on 2026-10-22 by under a month.

Google Cloud retirement date, now passed while the model still carries it on the lifecycle table rather than moving to the retired table. The Gemini API lists no shutdown date for this model, so the date binds Vertex AI deployments only. Google now names gemini-3.8-flash first and gemini-3.5-flash second. Google is inconsistent here: the Vertex AI generative AI release notes entry of 2026-04-02 gives 2026-10-16 for the whole Gemini 2.5 family, four days earlier, so plan against the earlier date. Separately, on 2026-09-18 the Gemini API limited 2.5 access to users who have used the models before, while stating they are not deprecated; Google points new projects to Gemini 3.5 Flash-Lite or 3.8 Flash.

Google Cloud retirement date, now passed but still listed on the lifecycle table; the Gemini API lists none. Google names gemini-3.8-flash, gemini-3.5-flash-lite or gemini-3.1-flash-lite. The Vertex AI generative AI release notes give 2026-10-16 for the same model, so plan against the earlier date. Separately, on 2026-09-18 the Gemini API limited 2.5 access to users who have used the models before, while stating they are not deprecated; Google points new projects to Gemini 3.5 Flash-Lite or 3.8 Flash.

Google Cloud retirement date, now passed but still listed on the lifecycle table; the Gemini API lists none. Google now names gemini-3.8-flash, gemini-3.5-flash-lite or Gemma 4, having dropped gemini-3.1-flash-lite from this row. The Vertex AI generative AI release notes give 2026-10-16 for the same model, so plan against the earlier date. Separately, on 2026-09-18 the Gemini API limited 2.5 access to users who have used the models before, while stating they are not deprecated; Google points new projects to Gemini 3.5 Flash-Lite or 3.8 Flash.

Google Cloud copy. The Claude API retired it on 2026-04-20.

Google Cloud copy, available to existing customers only during deprecation. The Claude API retired it on 2026-02-19.

Google Cloud copy, available to existing customers only during deprecation and with no replacement named on the partner deprecations page. The Claude API retired it on 2026-02-19 and Bedrock runs its regional copies to 2026-11-13.

The date moved out from 2026-10-03 to 2027-04-16, so the preview that looked like a four-month experiment now runs as long as Kimi-K2.6, which shares that date. Kimi-K2.5 goes first on 2027-01-26. No replacement is named for any of the three.

No replacement named. The OpenAI API put gpt-4.1 on notice on 2026-10-14 with a 2027-01-14 shutdown, three months earlier than Azure. Fine-tuned deployments run to 2027-10-14.

No replacement named. Fine-tuned deployments run to 2027-10-14.

The OpenAI API shut this one down on 2026-10-23, almost six months earlier, so an Azure deployment is the only place it keeps running into 2027. Fine-tuned deployments run to 2027-10-14.

The August 2024 snapshot outlives the May 2024 one by more than six months. The 2024-11-20 snapshot shares this retirement date but is only Legacy, not deprecated. Both snapshots are now on the OpenAI API notice of 2026-10-14, which shuts them down on 2027-01-14, three months before Azure. Fine-tuned deployments run to 2027-10-01.

No replacement named. Fine-tuned deployments run six months past the base model, to 2027-10-01. The OpenAI API put gpt-4o-mini-2024-07-18 on notice on 2026-10-14 with a 2027-01-14 shutdown, three months earlier than Azure.

Both v3 models are deprecated on 2026-11-15 and retire on 2027-02-15. Foundry names Cohere-embed-v5-fast and Cohere-embed-v5-pro; embed-v-4-0 itself runs to 2027-06-01. Changing embedding model means re-embedding the corpus.

Microsoft deprecated it on 2026-10-09 and points at a managed-compute Mistral Small 4 build. Mistral retired its own mistral-small-2503 endpoint on 2025-11-30, nearly 15 months earlier.

No replacement named. tts 001 and tts-hd 001 share the date. The OpenAI API runs whisper-1 to 2027-02-26 and tts-1 to 2027-01-06, so every one of these speech models dies earlier on Azure than on the OpenAI API.

The date moved out from 2026-10-01 to 2026-12-09 and the named replacement changed from gpt-5.1 to gpt-5.6-sol, so Azure now runs this snapshot about seven weeks past the OpenAI API shutdown of 2026-10-23. Standard deployments auto-upgrade at retirement; provisioned deployments do not and must migrate by hand.

Foundry names a managed-compute Ministral 4B build as the replacement, so the serverless path ends here. Mistral retired its own ministral-3b-2410 endpoint on 2025-12-31.

Mistral itself retired mistral-medium-2505 on 2026-08-31, so Foundry keeps serving a model that is already gone from la Plateforme for more than three further months.

Preview snapshot retiring well before the base DeepSeek-V4-Flash, which Foundry lists as GA to 2028-02-20 with this very snapshot named as its replacement. The two dates contradict each other, so plan against 2026-12-03.

Still GA on Foundry with no replacement named. Llama 4 Scout and Llama-3.3-70B-Instruct both run to 2027-01-31, so Maverick is the first of the surviving Llamas to go. Bedrock lists the same model to 2027-03-30.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Azure OpenAI date, which runs later than the OpenAI API shutdown. Deprecated on Azure means existing subscriptions keep deploying and new ones are blocked. Provisioned deployments are not auto-upgraded.

Foundry versions each snapshot of this rolling chat alias separately and retires them on their own dates: 2026-05-05 and 2026-06-24 on 2026-11-17, 2026-05-28 on 2026-12-08, 2026-08-06 on 2027-01-12 and 2026-10-07 on 2027-05-18. All are Preview and none names a replacement, so pinning a snapshot buys only months.

Both hosted versions, the Anthropic-infrastructure version 1 and the Azure-hosted version 2, carry this date. Anthropic's own table still shows the model Active with a tentative date of not sooner than 2026-10-15.

No replacement named. The OpenAI API retired codex-mini-latest on 2026-02-12.

Past its date, which matched the OpenAI API shutdown, yet Foundry still lists it as Preview rather than Retired and names no replacement. gpt-image-2 runs to 2027-10-21, the two gpt-image-2.5 models to 2027-09-09 and gpt-image-1-mini to 2027-04-07.

Foundry points users at the coherelabs-command-a-plus-05-2026 fp8, bf16 and w4a4 builds in the Hugging Face collection for managed compute deployment, so the serverless path becomes a self-managed one.

MAI-Image-2.5, MAI-Image-2.5-Flash and MAI-Image-2.5-Pro all retired on 2026-10-01 while still in preview, and the schedule names no replacement. They were themselves the named replacements for MAI-Image-2 and MAI-Image-2e, so that migration path is now a dead end.

Outlived DeepSeek-R1-0528 on Foundry by one month.

Azure also names mistral-ocr-4-0 as a replacement.

Mistral resells Z.ai models on la Plateforme and runs its own retirement schedule for them, separate from Z.ai's. GLM 5.3 is the named successor and is listed as a current open model there.

One day of notice between the deprecation and the retirement. OCR 3 (25.12) is still listed as a current premier model, so OCR 4.0 is the only one of the three that is gone.

The free Labs endpoint for Mistral's Lean 4 theorem prover retired with one day of notice and no replacement listed. The Apache-2.0 weights stay on Hugging Face, so self-hosting is now the only route.

Mistral Small 4 unifies instruct, reasoning, and coding, so reasoning workloads move there.

Coding agents move to the general Medium 3.5 model rather than a dedicated Devstral successor.

Three months of notice. Mistral Moderation 2 is the named successor and is listed as a current premier model.

The first free Labs endpoint for Mistral's Lean 4 theorem prover. Its named successor, Leanstral 1.5, then retired on 2026-09-30 with one day of notice and nothing to replace it, so the whole Labs theorem-prover line is gone.

The old flagship was retired into a Medium-tier replacement, not a Large one. Mistral Large 4 only reappeared as a public preview on 2026-10-06.

Retired with mistral-large-2411. The named replacement folds vision into the general Medium line rather than a dedicated Pixtral successor.

First GPT-5.1 family notice. Azure Foundry lists 2027-05-15 for gpt-5.1 2025-11-13 and still has it as GA, six weeks past the OpenAI API date.

The last Codex-branded model still on the API; gpt-5-codex through gpt-5.2-codex went on 2026-07-23. Azure Foundry lists 2027-08-24 for its copy.

gpt-5.4 and gpt-5.4-mini are not on this notice. Azure Foundry lists 2027-09-21 for its gpt-5.4-nano 2026-03-17 copy.

OpenAI names gpt-transcribe for file transcription and gpt-live-transcribe for streaming.

gpt-live-transcribe is the streaming alternative.

gpt-live-transcribe is the streaming alternative.

gpt-live-transcribe is the streaming alternative. Azure Foundry lists 2027-04-15 for its 2025-10-15 copy.

Family-level notice covering every gpt-realtime snapshot.

Family-level notice.

Family-level notice.

Family-level notice.

Family-level notice.

Family-level notice. OpenAI points mini users at the full gpt-audio-1.5, not a mini tier.

Family-level notice.

Family-level notice.

The named replacement belongs to the gpt-4o-mini-transcribe family, which itself retires on 2027-02-26.

Covers gpt-4o-2024-08-06 and gpt-4o-2024-11-20. The May 2024 snapshot is on the earlier 2026-04-22 notice and goes first. Azure Foundry lists 2027-04-14 for both snapshots, three months later than the OpenAI API.

First time gpt-4o-mini appears on an OpenAI API notice. Azure Foundry lists 2027-04-14 for the base model and 2027-10-01 for fine-tuned deployments.

The 2026-04-22 notice covered only gpt-4.1-nano, so this is the first date for gpt-4.1 itself. Azure Foundry also marks it deprecated but runs to 2027-04-14.

Azure Foundry marks it deprecated with a 2027-04-14 retirement, so an Azure deployment outlives the OpenAI API one by three months.

tts-1-hd is on the same notice. The named replacement is a realtime speech model, not a drop-in speech endpoint, so the migration is an API change and not just a model swap. Azure Foundry lists 2026-12-15 for its tts and tts-hd 001 copies, three weeks earlier.

Covers both snapshots, gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-tts-2025-12-15. With tts-1 and tts-1-hd on the same notice, this retires the whole dedicated text-to-speech endpoint family. Azure Foundry already retired the 2025-03-20 snapshot on 2026-10-15.

The replacement is gpt-5.6-sol with reasoning.mode set to pro.

The replacement is gpt-5.6-sol with reasoning.mode set to pro.

The notice now names gpt-image-2.5-sunburst or gpt-image-2.5-flare. Neither is a mini tier, so the cheapest image rate on the API rises when this one goes. Azure Foundry keeps its copy until 2027-04-07.

The notice now names gpt-image-2.5-sunburst or gpt-image-2.5-flare. Azure Foundry lists 2026-12-16 for its gpt-image-1.5 2025-12-16 copy, two weeks later.

The notice now names gpt-image-2.5-sunburst or gpt-image-2.5-flare.

Covers the alias, gpt-3.5-turbo-0125, and gpt-3.5-turbo-completions. Fine-tuned ft-gpt-3.5-turbo models retire the same day.

Covers the alias, gpt-4-0613, and both completions variants. Fine-tuned ft-gpt-4 models retire the same day.

Covers gpt-4-turbo-2024-04-09 and gpt-4-turbo-completions.

Covers gpt-4.1-nano-2025-04-14 and fine-tunes built on it. gpt-4.1 and gpt-4.1-mini are not on this notice.

Only the original May 2024 snapshot. The gpt-4o alias is not on the notice.

The notice now names gpt-image-2.5-sunburst or gpt-image-2.5-flare rather than gpt-image-2. Azure Foundry retires its gpt-image-1 2025-04-15 copy on the same date.

Covers the alias and o1-2024-12-17.

Covers o1-pro-2025-03-19. The replacement is gpt-5.6-sol with reasoning.mode set to pro.

Covers the alias and o3-mini-2025-01-31.

Covers o4-mini-2025-04-16 and fine-tuned ft-o4-mini-2025-04-16 models.

Announced with 20 days of notice and now in the past deprecations table. The notice itself names no model id, only the most capable cyber model available to your account, so confirm the target against your own model list before migrating.

The last instruct-style completion model on the API. Now in the past deprecations table.

Base model shut down 2026-09-28. Fine-tuned ft-babbage-002 models are on a separate notice and keep serving until 2026-10-23.

Base model shut down 2026-09-28. Fine-tuned ft-davinci-002 models are on a separate notice and keep serving until 2026-10-23.

The whole Videos API shut down the same day and OpenAI names no replacement. Snapshots sora-2-2025-10-06 and sora-2-2025-12-08 went with it. Azure Foundry kept its sora-2 2025-12-08 copy until 2026-10-15.

Includes sora-2-pro-2025-10-06. No replacement named; the Videos API itself is retired.

The alias and computer-use-preview-2025-03-11 were both removed.

The alias and o3-deep-research-2025-06-26 were both removed. Azure keeps its copy until 2026-11-19.

The alias and o4-mini-deep-research-2025-06-26 were both removed.

Limited to existing users since 2024, then formally retired in the September 2025 legacy batch alongside gpt-4-0125-preview. OpenAI named gpt-5 or gpt-4.1 as the target; the gpt-5 launch snapshot itself retires on 2026-12-11.

All three 32k variants (gpt-4-32k, gpt-4-32k-0314, gpt-4-32k-0613) shut down together.

gpt-3.5-turbo-16k-0613 shut down the same day.

Last of the original GPT-3 completion-API family. Replaced by chat-completions.

After retirement the slug is served by grok-imagine-image-2.0 at low quality, one cent per image cheaper at every resolution. grok-imagine-image (1.0) is not affected.

Slug now served by grok-4.3 at low reasoning effort and billed at grok-4.3 rates.

Slug now served by grok-4.3 with reasoning effort none and billed at grok-4.3 rates.

Slug now served by grok-4.3 at low reasoning effort and billed at grok-4.3 rates.

Slug now served by grok-4.3 with reasoning effort none and billed at grok-4.3 rates.

Slug now served by grok-4.3 at low reasoning effort. Set medium or high effort explicitly to match the old depth.

Slug now routes to grok-build-0.1.

Slug now served by grok-4.3 with reasoning effort none. Azure retired its grok-3 copy on 2026-05-01.

Redirects to grok-imagine-image-quality, which is itself retired on 2026-11-02.

Marked Deprecated in the voice pricing table on the xAI models page, with no announcement or retirement date given. grok-voice-3.0 is the only other voice model listed.