AI Model Shutdown Dates 2026-2027: Calendar and Migration Plan

Last updated: October 18, 2026

Between now and the end of June 2027, 192 dated AI model shutdowns are scheduled across the major model APIs, by the count in our model deprecations table. That count covers the official deprecation pages of OpenAI, Anthropic, Google (Gemini API and Vertex AI), Microsoft Foundry, Amazon Bedrock, Mistral and xAI, leaves out one managed agent that is not a model, and was re-checked against each provider's page on October 18, 2026. Four dates are likely to catch the most production code: GPT-4o and GPT-4.1 on the OpenAI API on January 14, 2027, OpenAI's realtime and audio families on January 20, whisper-1 and the gpt-4o transcribe models on February 26, and Claude Sonnet 4.5 on November 30, 2026.

The verdict: the date that applies to you depends on where you call the model, not on the model itself. The same Sonnet 4.5 retires on the Claude API on November 30, 2026 but runs on Amazon Bedrock until April 8, 2027. Gemini 2.5 Flash retires on Vertex AI on October 20, 2026, while the Gemini API has not announced a date at all. So inventory by platform first, then by model.

This post is the calendar plus a cross-platform migration checklist. For the OpenAI-only eval plan and replacement pricing, read Migrating Off Retired GPT Models in 2026; for the transcription shutdown, read the whisper-1 and gpt-4o-transcribe deep dive. The code samples below were type-checked against openai@7.38.0, @anthropic-ai/sdk@0.135.0 and @aws-sdk/client-bedrock@3.1152.0; no live API calls were made while writing this.

The six dates that matter most#

• October 20 is two days away. If you run Gemini 2.5 on Vertex AI, there is no time left for an eval cycle - treat it as a same-week fix.
• October 23 is the OpenAI legacy wave announced back on April 22, 2026. The GPT migration checklist linked above already walks through it, so it is not repeated here.
• November 30 is the only Claude API model with a scheduled retirement as of today. What changes in the replacement is covered in the Claude Sonnet 5.5 developer guide.
• January 14 covers the GPT-4o and GPT-4.1 snapshots many apps still pin, and it arrived with three months' notice (more on that below).
• January 20 retires the original realtime and audio families, so voice agents need a new model ID and a regression pass on latency and voice quality. The Realtime Voice API guide covers the API shape.
• February 26 is not a model swap for everyone: word timestamps, diarization and file translation need their own plan.

The full month-by-month calendar follows.

Month-by-month calendar#

Counts are rows in the /deprecations dataset with a shutdown date in that month (October counts only from October 19). A row is one model ID on one platform, so the same model on two clouds counts twice. Managed agents are excluded: the dataset also carries one Gemini API agent row on October 23, which is not counted or listed here.

October 2026 (42 shutdowns)#

• Vertex AI, Oct 20: gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite.
• Vertex AI, Oct 21: 16 open models served as Model-as-a-Service: deepseek-ocr-maas, deepseek-r1-0528-maas, deepseek-v3.1-maas, deepseek-v3.2-maas, glm-4.7-maas, glm-5-maas, gpt-oss-20b-maas, kimi-k2-thinking-maas, llama-3.3-70b-instruct-maas, minimax-m2-maas, multilingual-e5-large-instruct-maas, multilingual-e5-small-maas, qwen3-235b-a22b-instruct-2507-maas, qwen3-coder-480b-a35b-instruct-maas, qwen3-next-80b-a3b-instruct-maas and qwen3-next-80b-a3b-thinking-maas. Google names no hosted replacement; the page points to self-deploying the same weights from Model Garden.
• Gemini API, Oct 22: veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview and gemini-omni-flash-preview, replacement gemini-omni-1.1-flash.
• OpenAI, Oct 23: the legacy wave (17 IDs in the dataset, including ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano, ft-o4-mini, ft-babbage-002 and ft-davinci-002). o1-pro users move to gpt-5.6-sol with reasoning.mode: pro.
• Microsoft Foundry, Oct 23: the gpt-image-1 preview version.
• Mistral, Oct 30: zai-glm-5-2 (Z.ai GLM 5.2, hosted by Mistral), alternative zai-glm-5-3.

November 2026 (29 shutdowns)#

• xAI, Nov 2: grok-imagine-image-quality. This one redirects instead of failing (see below).
• Amazon Bedrock, Nov 13: Claude 3 Haiku, Claude 3 Sonnet, Claude 3.5 Sonnet v1 and v2, and Claude 3.7 Sonnet, but only in a handful of Regions (among them ap-northeast-2, ap-south-1, ap-southeast-1, eu-west-2, eu-central-2, sa-east-1 and ca-central-1). Claude 3 Haiku already reached end of life in most other Regions on September 10, 2026. For the other models, check modelLifecycle in each Region you use.
• Foundry, Nov 15: claude-haiku-4-5 and codex-mini.
• Gemini API, Nov 17: gemini-3.1-flash-live-preview and gemini-2.5-flash-native-audio-preview-12-2025 (replacement gemini-3.8-live); gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts (replacement gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts).
• Foundry, Nov 17: two gpt-chat-latest preview versions.
• Foundry, Nov 19: o1, o1-pro, o3, o3-pro and one other o3 variant (replacement gpt-5.6-sol); o3-mini and o4-mini (replacement gpt-5.6-terra).
• Vertex AI, Nov 19: gemini-3.6-flash, a short-term model, replacement gemini-3.8-flash.
• Foundry, Nov 24: claude-opus-4-5.
• Bedrock, Nov 26: AI21 Jamba 1.5 Large and Jamba 1.5 Mini.
• Anthropic and Foundry, Nov 30: claude-sonnet-4-5, replacement claude-sonnet-5-5.
• Bedrock, Nov 30: TwelveLabs Marengo Embed v2.7.
• OpenAI, Nov 30 (not models): the v1/prompts reusable prompts API, the Evals dashboard and API, and Agent Builder shut down. Evals go read-only on October 31.

December 2026 (21 shutdowns)#

• OpenAI, Dec 1: gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest, replacement gpt-image-2.5-sunburst or gpt-image-2.5-flare.
• Foundry, Dec 3: DeepSeek-V4-Flash-0731 (preview).
• Foundry, Dec 8: a gpt-chat-latest preview version (2026-05-28).
• Foundry, Dec 9: gpt-4o version 2024-05-13 (Standard deployments auto-upgrade to gpt-5.6-sol), plus Ministral-3B, mistral-medium-2505 and mistral-small-2503.
• OpenAI, Dec 11: gpt-5-2025-08-07 (to gpt-5.6-sol), gpt-5-mini-2025-08-07 (to gpt-5.6-terra), gpt-5-nano-2025-08-07 (to gpt-5.6-luna), o3-2025-04-16 (to gpt-5.6-sol), and gpt-5-pro-2025-10-06 and o3-pro-2025-06-10 (to gpt-5.6-sol with reasoning.mode: pro).
• Vertex AI, Dec 13: gemini-live-2.5-flash-native-audio.
• Foundry, Dec 15: whisper 001, tts 001, tts-hd 001 and Cohere-parse-v5.
• Foundry, Dec 16: gpt-image-1.5.

January 2027 (23 shutdowns)#

• OpenAI, Jan 6: tts-1, tts-1-hd, gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-tts-2025-12-15, replacement gpt-realtime-2.1-mini. The same day, existing fine-tuning customers can no longer create new fine-tuning jobs; inference on fine-tuned models continues until the base model is deprecated.
• Bedrock, Jan 8: Claude Opus 4.1 (us-east-1, us-east-2, us-west-2). It already retired on the Claude API on August 5, 2026.
• Foundry, Jan 12: a gpt-chat-latest preview version.
• OpenAI, Jan 14: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14, all to gpt-6-luna. Fine-tuned versions follow their base model.
• Foundry, Jan 15: claude-opus-5-5-fast.
• OpenAI, Jan 20: the gpt-realtime and gpt-4o-realtime families (to gpt-realtime-2.1), gpt-realtime-mini and gpt-4o-mini-realtime (to gpt-realtime-2.1-mini), gpt-audio, gpt-audio-mini, gpt-4o-audio and gpt-4o-mini-audio (to gpt-audio-1.5), and gpt-4o-mini-transcribe-2025-03-20 (to gpt-4o-mini-transcribe-2025-12-15).
• Foundry, Jan 26: Kimi-K2.5 (preview).
• Vertex AI, Jan 28: gemini-3.7-flash, replacement gemini-3.8-flash.

February 2027 (9 shutdowns)#

• Foundry, Feb 2: claude-opus-4-6.
• Foundry, Feb 9: gpt-5, gpt-5-mini, gpt-5-nano.
• Foundry, Feb 10: claude-sonnet-4-6.
• OpenAI, Feb 26: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize (replacement gpt-live-transcribe or gpt-transcribe), and /v1/audio/translations is replaced by v1/realtime/translations. The whisper-1 guide linked at the top covers the capability gaps.

March 2027 (31 shutdowns)#

• Foundry, Mar 2: gpt-audio and gpt-realtime version 2025-08-28.
• Bedrock, Mar 9 and Mar 21: Writer Palmyra X4 (writer.palmyra-x4-v1:0) and Palmyra X5 (writer.palmyra-x5-v1:0). These two appear on the Bedrock legacy page but are not yet in our dataset count.
• Gemini API and Vertex AI, Mar 15: gemini-2.5-flash-image, replacement gemini-3.1-flash-lite-image.
• Foundry, Mar 17: gpt-5-codex.
• Bedrock, Mar 30: the open-weight wave, 26 model IDs in our dataset: DeepSeek-R1 and DeepSeek-V3.1; Gemma 3 4B, 12B and 27B; Llama 3 8B and 70B, Llama 3.1 8B and 70B, Llama 3.3 70B, Llama 4 Maverick and Scout; MiniMax M2 and M2.1; Devstral 2 123B, Mistral 7B, Mixtral 8x7B, Voxtral Mini 3B and Voxtral Small 24B; Kimi K2 Thinking; Nemotron Nano 9B v2 and 12B v2; Qwen3-Coder 30B and 480B; GLM 4.7 and GLM 4.7 Flash.

April 2027 (23 shutdowns)#

• OpenAI, Apr 1: gpt-5.1 and gpt-5.3-codex (to gpt-6-sol), gpt-5.4-nano (to gpt-6-luna). Announced October 1 with six months' notice. Our August checklist said gpt-5.1 and newer base models had no deadline; OpenAI set April 1, 2027 on October 1, and that checklist has now been corrected.
• Vertex AI, Apr 1: text-embedding-004, text-embedding-005, text-multilingual-embedding-002 and multimodalembedding@001. Embedding changes mean re-indexing, so start early.
• Foundry, Apr 2: claude-mythos-preview.
• Foundry, Apr 6: claude-opus-4-7, gpt-audio-mini 2025-10-06, and the grok-4-20-reasoning and grok-4-20-non-reasoning previews.
• Foundry, Apr 7: gpt-5-pro and gpt-image-1-mini.
• Bedrock, Apr 8: Claude Sonnet 4.5, across 31 commercial Regions.
• Foundry, Apr 14: gpt-4o 2024-08-06 and 2024-11-20 (auto-upgrade target gpt-5.1), gpt-4o-mini, gpt-4.1, gpt-4.1-mini and gpt-4.1-nano.
• Foundry, Apr 15: gpt-4o-transcribe-diarize.
• Foundry, Apr 16: Kimi-K2.6 (preview).

May 2027 (7 shutdowns)#

• Gemini API, May 7: gemini-3.1-flash-lite, replacement gemini-3.5-flash-lite.
• Foundry, May 15 and May 18: gpt-5.1, gpt-5.1-codex and gpt-5.1-codex-mini (May 15), gpt-5.1-codex-max and a gpt-chat-latest preview (May 18).
• Foundry, May 20: model-router version 2025-11-18.

June 2027 (7 shutdowns)#

• Foundry, Jun 8: gpt-5.2.
• Foundry, Jun 15: gpt-4o-mini-transcribe, gpt-4o-mini-tts and gpt-audio-mini version 2025-12-15.
• Foundry, Jun 25: gpt-realtime-2.1 and gpt-realtime-2.1-mini.
• Foundry, Jun 30: claude-sonnet-5.

One data-quality caveat: Microsoft's retirement schedule currently lists gpt-realtime-mini twice per version, with conflicting dates (2025-10-06 shows both April 6, 2027 and September 21, 2026; 2025-12-15 shows both June 15, 2027 and December 15, 2026). We do not pick one. If you deploy it, confirm the date for your deployment in the Foundry portal or the Models API.

Same model, different clocks#

This is the part most migration plans miss. A model ID retires per platform, and the platforms run on different policies.

Why the clocks differ comes straight from each provider's policy page:

• Anthropic's model deprecations page says its dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry, while Amazon Bedrock and Google Cloud set their own schedules. Even so, Microsoft's schedule gives hard Foundry dates for Haiku 4.5 and Opus 4.5 where Anthropic's page only lists floors, and for Opus 4.6, Sonnet 4.6 and Opus 4.7 the Foundry dates (Feb 2, Feb 10 and Apr 6, 2027) land before Anthropic's not-sooner-than floors (Feb 5, Feb 17 and Apr 16, 2027). A Foundry user planning against the Anthropic floor would miss the deadline. Where the two pages differ, plan against the platform you actually call, and take the earlier date.
• Bedrock guarantees models launched before September 7, 2026 at least 12 months on the platform and at least 6 months in Legacy, which is why it tends to run later than the first-party date.
• Foundry sets a GA retirement date at launch: 18 months, or 12 months for Anthropic, DeepSeek, Fireworks and Mistral AI models. That fixed clock is why Foundry's Opus 4.5 date is firm while Anthropic only publishes a floor.
• Vertex AI separates long-term models (at least 12 months after release) from short-term ones with a fixed date and at least 45 days to migrate.

%% title: One model, three retirement clocks
flowchart LR
    M[claude-sonnet-4-5] --> A[Claude API]
    M --> F[Microsoft Foundry]
    M --> B[Amazon Bedrock]
    A --> A1[Retires Nov 30 2026]
    F --> F1[Retires Nov 30 2026]
    B --> B1[Legacy Oct 8 2026]
    B1 --> B2[Extended access Jan 8 2027]
    B2 --> B3[EOL Apr 8 2027]

The practical upshot: a later cloud date is a buffer, not a plan. Bedrock says to expect higher, provider-set pricing during extended access, and Foundry's auto-upgrade can change your model without a deploy. If you run the same model on two platforms (a common failover setup), the earlier date is your real deadline, because your fallback disappears first.

What happens on the day: fail, 410, or redirect#

Providers do not all fail the same way, and the quiet failures are the dangerous ones.

• Request fails: OpenAI, Anthropic, the Gemini API ("completely turned off"), Vertex AI, Bedrock and Groq. Retired IDs return an error.
• 404: Mistral returns 404 for retired identifiers.
• 410 Gone or silent upgrade: Foundry returns 410 Gone for retired models, unless a Standard deployment is set to auto-upgrade. Then the deployment moves to the replacement (for example gpt-4o to gpt-5.1 or gpt-5.6-sol) and keeps answering. Provisioned deployments are never auto-upgraded.
• Redirect: xAI keeps the old slug resolving. After November 2, grok-imagine-image-quality requests are served by grok-imagine-image-2.0 with quality set to low and billed at that model's rate, which xAI says is $0.01 per image cheaper at every resolution; grok-imagine-image-pro, which already redirects, follows.
• Auto-route: on the Gemini API, gemini-3.5-flash and gemini-3.7-flash are not shutting down; their endpoints route to gemini-3.6-flash and gemini-3.8-flash.

%% title: What a request to a retired model returns
flowchart TD
    R[Request names retired model] --> P{Which platform}
    P -->|OpenAI Anthropic Gemini Bedrock| E[Error - request fails]
    P -->|Mistral| N[404]
    P -->|Foundry Standard| U{Auto-upgrade on}
    U -->|yes| S[Served by replacement]
    U -->|NoAutoUpgrade| G[410 Gone]
    P -->|Foundry Provisioned| G
    P -->|xAI| X[Served by replacement at its rate]

A redirect looks like nothing broke. Your error rate stays flat while the serving model, output quality and latency change underneath you; in xAI's case the price goes down, but you still get a different model than the one you tested. xAI documents that the response model field reports the model that actually served the request, so log it and alert when it stops matching the model you asked for. The same check is worth adding on Foundry and the Gemini API, where an auto-upgrade or alias reroute can change the serving model; confirm what each platform reports in that field before you rely on it.

Notice policies compared#

Credit where it is due: Vertex AI states that listed retirement dates may be extended but will not move earlier, which is the most planning-friendly promise on this list. OpenAI exposes the shutdown date in the API itself, and Anthropic and Bedrock go further with full lifecycle fields.

One neutral observation: OpenAI's October 14 notice for GPT-4o and GPT-4.1 gives three months (October 14 to January 14), while the same page states a six-month minimum for generally available models unless safety or compliance require a faster timeline. The notice does not give a reason. The lesson for planning is general: work against the published date, not the policy floor. OpenAI also says some developers may be able to provision dedicated capacity for continued access after a shutdown date by contacting sales, which is worth knowing if January is unworkable.

Two more terms trip people up:

• Anthropic "Not sooner than" dates are floors on Active models, not shutdowns. claude-haiku-4-5-20251001 had a floor of October 15, 2026; that date has passed and the model is still Active on the Claude API. The real deadline arrives with a deprecation notice and at least 60 days' warning. Other floors: claude-opus-4-6 February 5, 2027, claude-sonnet-4-6 February 17, 2027, claude-opus-4-7 April 16, 2027, claude-opus-4-8 May 28, 2027, claude-sonnet-5 June 30, 2027. On Foundry, the first three retire earlier than these floors (see the clocks table above).
• Foundry's API and docs use "Deprecated" differently. In the docs, Deprecated means existing customers only. In the Models API, the value Deprecating corresponds to that docs stage, and Deprecated means retired, returning 410. Read the API value, not the word.

Migration checklist#

Step 1: inventory by platform and model#

Start with code and config. This grep pulls the OpenAI, Anthropic and Google IDs on this calendar out of a repo, from the October 23 legacy wave through the April 1 OpenAI wave (gpt-5.1, gpt-5.3-codex, gpt-5.4-nano), and counts each one:

grep -rhoE "gpt-4o[a-z0-9.-]*|gpt-4\.1[a-z0-9.-]*|gpt-4(-[a-z0-9-]+)?\b|gpt-3\.5[a-z0-9.-]*|gpt-5-(mini-|nano-|pro-)?20[0-9-]+|gpt-5\.1[a-z0-9-]*|gpt-5\.3-codex|gpt-5\.4-nano|\bo(1|3|4)(-mini|-pro)?\b|tts-1(-hd)?|whisper-1|gpt-realtime[a-z0-9.-]*|gpt-audio[a-z0-9.-]*|gpt-image-1[a-z0-9.-]*|chatgpt-image-latest|claude-sonnet-4-5[a-z0-9-]*|claude-opus-4-1[a-z0-9-]*|claude-3[a-z0-9.-]*|gemini-2\.5[a-z0-9.-]*|text-embedding-00[45]" \
  --exclude-dir={node_modules,.git,dist,build,.next} . \
  | sed -E 's/[.-]+$//' \
  | grep -vxE 'gpt-realtime-2\.1(-mini)?|gpt-audio-1\.5' \
  | sort | uniq -c | sort -rn

What each stage does: the first grep matches the retiring families, including the bare gpt-4 alias and Bedrock-style IDs such as us.anthropic.claude-opus-4-1-20250805-v1:0 (it reports the claude-opus-4-1-... part). The sed trims a trailing dot or dash picked up from prose. The second grep drops OpenAI's recommended replacements, gpt-realtime-2.1, gpt-realtime-2.1-mini and gpt-audio-1.5, which the family patterns would otherwise count. We ran it under GNU grep 3.11 against a file of 30 retiring IDs (among them gpt-4o-2024-08-06, gpt-4o-mini, claude-sonnet-4-5, gemini-2.5-flash-lite, gpt-4, gpt-4-1106-preview, gpt-5.3-codex, gpt-5.4-nano and gpt-4o-mini-transcribe-2025-12-15) and 23 near-misses (gpt-realtime-2.1, gpt-audio-1.5, gpt-5.6-sol, gpt-6-luna, gpt-image-2.5-sunburst, gpt-5.4, claude-sonnet-4-6, claude-sonnet-5-5, gemini-3.5-flash-lite and others): it reported all 30 and none of the near-misses.

Three caveats. gpt-4o-mini-transcribe-2025-12-15 stays in the output on purpose: it is the January 20 target for the older snapshot, but the whole gpt-4o-mini-transcribe family shuts down on February 26. gpt-realtime-2.1 is filtered out because it is OpenAI's replacement, yet it retires on Foundry on June 25, 2027, so Foundry users should drop that exclusion. And gemini-2.5-* matches carry the Vertex AI date; the Gemini API has not announced one for the base 2.5 models. Bedrock and Foundry deployments are better inventoried through their lifecycle APIs in Step 2 than by grep.

Drop -h and -o and add -n to see where each one lives. Then check what actually receives traffic, because a grep finds dead code and misses IDs stored in a database, a feature-flag service or a secret store:

• OpenAI: the usage dashboard, filtered by model.
• Anthropic: Claude Console, Usage, Export, which gives a CSV broken down by API key and model.
• Bedrock: invocation metrics per model ID, per Region.
• Foundry: the deployments list, including each deployment's model version and upgrade setting.

Write the result as a table of platform, Region, model ID, date and owner. Sort by date.

Step 2: ask each API#

Four providers expose lifecycle data you can poll: three through their SDKs, plus a REST call for Foundry. Run the TypeScript snippets as ES modules (they use top-level await).

OpenAI returns an optional shutdown_date (a YYYY-MM-DD string or null) on GET /v1/models and on the retrieve endpoint:

// List every OpenAI model your key can see that has an announced shutdown date.
import OpenAI from "openai";

const client = new OpenAI();

for await (const model of client.models.list()) {
  if (model.shutdown_date) {
    console.log(`${model.id}\t${model.shutdown_date}`);
  }
}

Anthropic's Models API returns lifecycle (active, deprecated or retired) plus deprecated_at and retires_at, which are null while a model is active or unscheduled:

// List Claude models that are deprecated or retired, with their scheduled retirement.
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

for await (const model of client.models.list()) {
  if (model.lifecycle !== "active") {
    console.log(`${model.id}\t${model.lifecycle}\tretires_at=${model.retires_at ?? "not scheduled"}`);
  }
}

By default the list returns active and deprecated models; retired ones only appear when you ask for them with the lifecycle filter, for example client.models.list({ lifecycle: ["deprecated", "retired"] }). Anthropic's docs also note that a past retires_at on a deprecated model means retirement is overdue, so treat lifecycle as the signal.

Bedrock lifecycle differs by Region, so run this once per Region you use:

// List Bedrock foundation models in this Region that are in the LEGACY state.
import { BedrockClient, ListFoundationModelsCommand } from "@aws-sdk/client-bedrock";

const client = new BedrockClient({ region: "us-east-1" });
const { modelSummaries = [] } = await client.send(new ListFoundationModelsCommand({}));

for (const m of modelSummaries) {
  const life = m.modelLifecycle;
  if (life?.status === "LEGACY") {
    console.log(
      `${m.modelId}\tEOL=${life.endOfLifeTime?.toISOString().slice(0, 10) ?? "unknown"}` +
        `\textendedAccess=${life.publicExtendedAccessTime?.toISOString().slice(0, 10) ?? "n/a"}`,
    );
  }
}

For Foundry, Microsoft documents a subscription-scoped Models API that lists every model in a Region:

GET https://management.azure.com/subscriptions/{subscriptionId}/providers/Microsoft.CognitiveServices/locations/{location}/models?api-version=2024-10-01

Read lifecycleStatus, deprecation.inference, deprecation.fineTune and the per-SKU deprecationDate. Remember the mapping: Deprecating still serves, Deprecated is retired. Foundry also posts health advisories to Azure Service Health, where you can create an alert rule for email or webhook notifications.

The Gemini API, Vertex AI, Mistral and xAI do not document a lifecycle field. Watch their pages, or use the /deprecations table, which records the date each provider page was last verified.

Step 3: map to the named replacement, and test one alternative#

Every provider names a recommended replacement, but treat it as a starting point, not an equivalent. Evaluate at least one alternative alongside it. On Foundry this matters more: Microsoft selects the official replacement roughly 90 to 120 days before retirement, so for dates in 2027 you may not have a named target yet.

Step 4: evaluate before you switch#

Build a golden set from real production requests and compare quality, latency and token cost on the old model, the named replacement and your alternative. The GPT migration checklist linked at the top has a five-day version of this plan that works for any provider.

Step 5: cut over behind config#

Keep the model ID in config, not code, so a rollback is a config change. Ship to a slice of traffic first. Two platform-specific traps:

• Foundry: set versionUpgradeOption on purpose. OnceNewDefaultVersionAvailable upgrades when a new default ships, OnceCurrentVersionExpired upgrades only at retirement, and NoAutoUpgrade stops the deployment at retirement. A loud 410 is easier to debug than a silent model change.
• Bedrock: in Legacy, existing customers may lose access after 15 days of inactivity. A fallback path you never exercise can quietly disappear before end of life, so send it periodic traffic or drop it from the plan.

Step 6: make it a weekly job#

Schedule the Step 2 scripts weekly, diff the output against last week, and alert on any new date. Add the response model field check from the redirect section. Shutdown notices now arrive with as little as three months' warning on GA models and two weeks on previews, so a once-a-quarter review is too slow.

%% title: Shutdown migration loop
flowchart LR
    I[Inventory IDs by platform] --> Q[Query lifecycle APIs]
    Q --> D[Sort by earliest date]
    D --> E[Eval replacement plus one alternative]
    E --> C[Flagged cutover]
    C --> W[Weekly lifecycle diff]
    W --> Q

Special cases#

• Embeddings (Vertex AI, April 1, 2027): a new embedding model produces vectors in a different space, so this is a full re-index of your corpus, not a model swap. Budget the compute and run old and new indexes side by side until recall checks out.
• Audio and realtime (OpenAI, January 6, January 20 and February 26): the TTS replacement is a realtime model, so moving off tts-1 changes the API shape as well as the ID. See the Realtime Voice API guide and the whisper-1 guide linked above.
• Fine-tunes on Azure outlive their base model: gpt-4o 2024-08-06 and gpt-4o-mini fine-tune deployments retire October 1, 2027, and the gpt-4.1 family fine-tunes October 14, 2027. On OpenAI, fine-tuned models follow the base model's date.

Providers with nothing (or almost nothing) scheduled#

• Groq: no shutdowns after September 21, 2026.
• DeepSeek and Cohere: no dated shutdowns in the window.
• xAI: only the November 2 image redirect.
• Mistral: only zai-glm-5-2 on October 30.
• Anthropic (Claude API): only Sonnet 4.5 on November 30; claude-mythos-preview is deprecated with retirement "To be announced".

A fairness note: OpenAI, Bedrock and Foundry dominate this calendar partly because they publish the most detail, including every snapshot, Region and preview version. A longer list is not a worse platform, and a short one does not mean nothing will retire. It may just mean the notice has not been written yet.

FAQ#

When does GPT-4o shut down?#

On the OpenAI API, gpt-4o-2024-05-13 shuts down on October 23, 2026, and gpt-4o-2024-08-06, gpt-4o-2024-11-20 and gpt-4o-mini-2024-07-18 shut down on January 14, 2027, with gpt-6-luna as the recommended replacement. On Microsoft Foundry the dates are later: December 9, 2026 for 2024-05-13 and April 14, 2027 for the other versions.

Is Claude Sonnet 4.5 being retired?#

Yes. claude-sonnet-4-5-20250929 retires on the Claude API and Microsoft Foundry on November 30, 2026, and Anthropic recommends claude-sonnet-5-5. On Amazon Bedrock it reaches end of life on April 8, 2027, with public extended access from January 8, 2027, during which Bedrock says to expect higher, provider-set pricing.

What happens if I call a model after its shutdown date?#

On most platforms the request fails: Foundry returns 410 Gone and Mistral returns 404. xAI is the exception, redirecting the old slug to a replacement and billing at that model's rate, which for grok-imagine-image-quality xAI says is $0.01 per image cheaper. Foundry Standard deployments may also auto-upgrade to the replacement instead of failing, depending on versionUpgradeOption.

How can I check shutdown dates programmatically?#

OpenAI returns shutdown_date on GET /v1/models. Anthropic's Models API returns lifecycle and retires_at. Bedrock returns modelLifecycle from ListFoundationModels, per Region. Foundry's Models API returns lifecycleStatus and deprecation.inference. The Gemini API, Vertex AI, Mistral and xAI document no equivalent field.

Is Gemini 2.5 Flash shutting down?#

On Vertex AI, yes: gemini-2.5-flash, gemini-2.5-pro and gemini-2.5-flash-lite retire on October 20, 2026. The Gemini API lists no shutdown date for gemini-2.5-flash as of October 18, 2026.

Sources#

• OpenAI API deprecations
• Anthropic model deprecations
• Anthropic Models API: list models
• Gemini API deprecations
• Gemini API models
• Vertex AI (Gemini Enterprise Agent Platform) model versions and lifecycle
• Vertex AI open model deprecations
• Microsoft Foundry models lifecycle and support policy
• Microsoft Foundry model retirement schedule
• Amazon Bedrock model lifecycle
• Amazon Bedrock model lifecycle (Legacy)
• Mistral model lifecycle
• Mistral models
• xAI: grok-imagine-image-quality migration (Nov 2)
• Groq deprecations
• openai on npm
• @anthropic-ai/sdk on npm
• @aws-sdk/client-bedrock on npm

Continue Reading#

• Migrating Off Retired GPT Models in 2026: A Working Checklist - the OpenAI eval-before-switch plan and replacement pricing this calendar points to.
• Migrate Off whisper-1 and gpt-4o-transcribe Before Feb 26, 2027 - the capability gap check for the February transcription shutdown.
• Claude Sonnet 5.5 Developer Guide: Pricing and API Changes - what changes when you move off Sonnet 4.5 before November 30.