LLM model deprecation policy comparison 2027: 45 days to 18 months

Summary. Every major AI platform now publishes a model retirement policy, and no two match. OpenAI promises at least six months' notice for generally available models, three months for specialised variants and as little as two weeks for previews, yet gave GPT-4o and GPT-4.1 users 92 days in October 2026. Anthropic guarantees at least 60 days' notice and lists a not-sooner-than date for every active Claude model, such as 7 October 2027 for Claude Haiku 5.5. Google's Gemini API publishes only earliest shutdown dates and reroutes two old endpoints to newer models, while the Gemini Enterprise Agent Platform (the renamed Vertex AI) promises most models at least 12 months and gives short-term models 45 days. Amazon Bedrock applies a new policy to models launched on or after 7 September 2026: each model card shows an earliest end-of-life (EOL) date and a Legacy period of six months or 45 days. Microsoft Foundry retires GA models 18 months after launch, or 12 months for Anthropic, DeepSeek, Fireworks and Mistral models, and auto-upgrades Standard deployments. The same model can die on different days: Claude Sonnet 4.5 leaves the Claude API on 30 November 2026 but stays on Bedrock until 8 April 2027. Plan for one forced migration per model per year, pin dated versions and feed each vendor's lifecycle signal into monitoring.

About this article. This comparison is for CTOs and platform engineers who run production workloads across LLM providers. Notice periods, dates and lifecycle fields were checked on 1 November 2026 against OpenAI, Anthropic, Google, AWS and Microsoft documentation. Vendors revise these pages often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What changed in 2026, and what retires before April 2027

Four shifts in 2026 matter for 2027 budgets:

• Bedrock split its policy. Models launched on or after 7 September 2026 follow a per-model-card policy. Older models stay on the legacy policy, which includes a public extended access phase at prices the model provider sets.

• Google renamed and tiered. Vertex AI documentation now sits under Gemini Enterprise Agent Platform, which separates models with a 12-month floor from short-term models with 45 days' notice.

• OpenAI cleared its catalogue. It shut down a dozen snapshots on 23 October 2026, including gpt-4o-2024-05-13, o1, o4-mini and gpt-4.1-nano, and on 14 October gave the remaining GPT-4 series three months rather than six.

• Foundry runs shorter partner lifecycles. Microsoft gives Anthropic, DeepSeek, Fireworks and Mistral models 12 months, and SpaceXAI Grok models from 4.7 onwards a six-month minimum.

The near-term calendar below is compiled from each vendor's page and is a selection, not the full list.

Two rows need care. Foundry now lists gpt-5-mini on 13 July 2027, not with gpt-5 and gpt-5-nano on 9 February; our Azure OpenAI GPT-5 retirement guide covers that move. And Anthropic's own floors for the two 4.6 models are later than Foundry's dates, as our Claude Opus 4.6 and Sonnet 4.6 retirement guide explains.

OpenAI: six months on paper

OpenAI sets three minimum notice periods, which it can shorten for safety or compliance reasons:

• Generally available models: at least six months.

• Specialised variants such as chat, Codex and deep research variants: at least three months.

• Preview models with preview in the name: much shorter, with two weeks as OpenAI's example.

Legacy means no longer updated; deprecated means a shutdown date exists. Fine-tuned models in recent notices follow their base model. After shutdown, some developers may be able to arrange dedicated capacity through sales, which is a negotiation rather than a policy.

The 1 October 2026 notice for gpt-5.1, gpt-5.3-codex and gpt-5.4-nano gave exactly 182 days. The 14 October 2026 notice for GPT-4o and GPT-4.1 gave 92 days, about half the GA promise. Our GPT-4o and GPT-4.1 shutdown guide covers that migration. Audio models run on separate dates, 6 January, 20 January and 26 February 2027, covered in our tts-1 and Realtime shutdown guide.

The engineering hook is shutdown_date. GET /v1/models returns it as a YYYY-MM-DD string, or null when no date is set, so a weekly job can flag any model ID your code still calls.

Anthropic: 60 days, plus a published floor

Anthropic promises at least 60 days' notice before retiring a publicly released model. Recent history shows it uses close to the minimum:

The offset is a published floor. Every active model carries a tentative not-sooner-than retirement date: 22 September 2027 for claude-opus-5-5, 28 September 2027 for claude-sonnet-5-5 and 7 October 2027 for claude-haiku-5-5. The Models API now reports lifecycle and retires_at for each model; retires_at stays null while a model is active, so alert on any change.

Three caveats apply. First, Anthropic's dates cover the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry; Bedrock and the Google-managed offering set their own schedules. Second, Foundry does not always agree. Anthropic's overview says Foundry follows its schedule, yet Microsoft lists claude-haiku-4-5 for retirement on 15 November 2026 while Anthropic still shows the model as Active with a floor of 15 October 2026, so trust the platform you call and its own notices. Third, retirements are not only about models. API version 2026-10-21 returns a 400 error for temperature, top_p and top_k on every model, and the SDK releases of 21 October 2026, such as Python 2.0.0 and TypeScript 1.0.0, send that version by default. A dependency upgrade can break code that never changed its model string.

Google: two products, two rulebooks

Gemini API (Google AI Studio)

Google's deprecations page lists only the earliest dates a model might be retired, with exact dates sent later. Its models page defines four version types:

• Stable (gemini-3.6-flash): usually unchanged, and recommended for production.

• Preview: typically billable, with tighter rate limits and at least two weeks' deprecation notice.

• Latest (gemini-flash-latest): swapped with every release, with two weeks' email notice for breaking changes.

• Experimental: no availability promise.

Two behaviours stand out. Gemini 2.5 Pro, Flash and Flash-Lite are not deprecated but are limited to users who have used them before. And retired endpoints can stay alive as redirects: since 8 October 2026, requests to gemini-3.7-flash route to gemini-3.8-flash, and requests to gemini-3.5-flash route to gemini-3.6-flash. Your code keeps working on a different model, at that model's price, and both targets double in price on 1 January 2027.

Gemini Enterprise Agent Platform (formerly Vertex AI)

The enterprise platform commits more firmly. Its lifecycle page, updated on 30 October 2026, lists models available for at least 12 months after release and says retirement dates can be extended but will not move earlier. Short-term models, such as gemini-3.6-flash, gemini-3.7-flash and the cyber variants, stay active until a replacement exists, then get a fixed date with at least 45 days to migrate.

So gemini-3.8-flash (released 2 September 2026) shows 2 September 2027 or later, while gemini-3.6-flash retires on 19 November 2026. Gemini 2.5 Pro, Flash and Flash-Lite retired on the Agent Platform on 20 October 2026, even though the Gemini API still serves them to existing users. Google's page treats a retired model as an error to fix by updating the model ID, and describes no reroute.

Google-managed Claude follows Google's partner model deprecations page, which has mostly used six-month windows: Claude 3.7 Sonnet was deprecated on 11 November 2025 and shut down on 11 May 2026, 181 days later. Claude 3 Opus got 32 days in 2025, so the pattern is not a promise.

Amazon Bedrock: a new policy from 7 September 2026

Bedrock's lifecycle page covers models launched on or after 7 September 2026. Each model card shows two items:

1. EOL no sooner than: the model will not reach end of life before this date.

1. Legacy period: the notice period before EOL, either six months or 45 days. AWS says most models get six months.

Claude Haiku 5.5's card is typical: launched 7 October 2026, EOL no sooner than 7 October 2027, Legacy period six months. GPT-6.1 Sol is the exception AWS names. It follows OpenAI's first-party terms, including OpenAI's six-month notice for GA models, with no separate Bedrock minimum.

Legacy bites before EOL. New customers cannot adopt the model, existing customers may lose access after 15 days of inactivity, and you cannot create new Provisioned Throughput or new fine-tuning jobs on it. Existing custom-model deployments keep running. At EOL the model is removed from all Regions unless you have a private arrangement with the provider, and nothing migrates automatically.

Models launched before 7 September 2026 keep the legacy rules: at least 12 months on Bedrock, at least six months in Legacy, and, for EOL dates after 1 February 2026, a public extended access phase after at least three months in Legacy, with higher pricing set by the provider. Claude Sonnet 4.5 shows the pattern: Legacy on 8 October 2026, extended access from 8 January 2027, EOL on 8 April 2027.

Regional rows can run shorter. In ap-south-1, the five Claude models in our calendar show a Legacy date of 10 October 2026 and an EOL date of 13 November 2026, 34 days apart. Lifecycle state also differs by Region, so read the modelLifecycle field in each Region you use.

Microsoft Foundry: 18 months, auto-upgrade by default

A standard GA model's retirement date is set programmatically at launch, 18 months out:

The gpt-5.6 models show the arithmetic: version 2026-07-09 retires on 11 January 2028, 551 days later. The schedule does not always follow the rule, though. gpt-5.1 (2025-11-13) and gpt-5-mini (2025-08-07) both show 13 July 2027, so read the date rather than computing it. Notice emails go out at least 60 days before a GA retirement and 30 days before a preview retirement. Preview deployments are force-upgraded or terminated. Microsoft says retirement dates cannot be extended, and it reserves an emergency retirement for security or compliance problems.

Global Standard, Data Zone Standard and Standard deployments auto-upgrade to the named replacement at retirement, controlled by versionUpgradeOption (OnceNewDefaultVersionAvailable, OnceCurrentVersionExpired or NoAutoUpgrade). Provisioned deployments never auto-upgrade, and Batch needs a side-by-side move. Partner models differ: Anthropic, DeepSeek, Fireworks and Mistral get 12 months. Fireworks models on per-token offerings carry only a 15-day notice period.

One trap: the Models API reports the docs' Deprecated stage as lifecycleStatus value Deprecating, and uses Deprecated to mean retired. Alert on deprecation.inference dates, not the label.

Head-to-head comparison

The table summarises the policy pages cited above.

Same model, different retirement dates

The platform you call decides when your model dies.

Hyperscalers usually keep third-party models longer, but not always. A multi-cloud routing layer, such as the gateways in our LiteLLM, Portkey, Kong and TrueFoundry comparison, turns these gaps into a fallback plan.

Decision table: which lifecycle fits your workload

Worked example: a three-model stack on 1 November 2026

A team runs GPT-4.1 on OpenAI for extraction, Claude Sonnet 4.5 on Bedrock in Mumbai (ap-south-1) for an assistant, and gemini-3.6-flash on the Agent Platform for summaries. Assume, as our own estimate, that each migration needs 42 days: two weeks each for evals, prompt fixes and a canary.

The Gemini workload is already late: switch the model ID to gemini-3.8-flash this week and run evals afterwards. On the Gemini API the same model has no shutdown date announced, so the platform choice alone created the emergency.

Adoption timing matters too. On Foundry, a model adopted six months after launch leaves 12 months of use (18 minus 6). On the Agent Platform, the same lag can leave as little as six months (12 minus 6). Adopt new models early in their life, or budget two migrations a year per workload.

India, EU and US data notes

India. Bedrock lifecycle is per Region. In ap-south-1, five older Claude models reach EOL on 13 November 2026, and Llama 3 70B and 8B on 30 March 2027. For the likely replacements, AWS's cards for Claude Haiku 5.5, Opus 5.5 and Sonnet 5.5 mark Mumbai and Hyderabad as Global cross-Region only, with no in-Region or Geo option, and AWS says Global routing can reach any Region worldwide. That can change where prompts are processed, so re-run any DPDP assessment.

EU. On Foundry, new models arrive in Global Standard first, then Global Provisioned, then Data Zone, then regional deployments. An EU Data Zone customer can find the named replacement testable globally but not yet in-zone, so ask Microsoft for Data Zone timing early.

US government. Azure Government clouds usually support one version of a model at a time, with a 30-day overlap when a new version arrives.

Implementation checklist

1. Inventory every model ID in code, configs and no-code tools, including aliases.

1. Pin dated snapshots in production.

1. Automate the sweep. Run a weekly job against each lifecycle signal and open a ticket for any date inside 180 days.

1. Set Foundry upgrade policy deliberately with versionUpgradeOption, and subscribe to Azure Service Health under Azure OpenAI Service, the name Microsoft uses for these alerts.

1. Record platform-specific dates, not provider dates. AWS says only the dates on its model cards apply to Bedrock usage.

1. Keep a golden set per workflow so any migration starts with evals.

1. Route through a gateway, with a second provider as fallback.

1. Negotiate lifecycle terms in enterprise contracts.

A minimal weekly sweep for OpenAI and Bedrock:

            import boto3
from openai import OpenAI

IN_USE = {"gpt-4.1-2025-04-14", "anthropic.claude-sonnet-4-5-20250929-v1:0"}

for m in OpenAI().models.list():
    if m.id in IN_USE and getattr(m, "shutdown_date", None):
        print(f"OpenAI {m.id} shuts down {m.shutdown_date}")

for region in ["ap-south-1", "us-east-1"]:
    bedrock = boto3.client("bedrock", region_name=region)
    for m in bedrock.list_foundation_models()["modelSummaries"]:
        status = m.get("modelLifecycle", {}).get("status")
        if m["modelId"] in IN_USE and status != "ACTIVE":
            print(f"Bedrock {region} {m['modelId']} is {status}")
          

We ran this against a mocked OpenAI endpoint and a stubbed Bedrock client, not live accounts. Pipe the output into your alerting.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our LLM migration and cost optimisation service inventories the models your systems call across vendors, maps each one to its platform-specific date, builds golden-set evals and wires lifecycle checks into your monitoring. Talk to us at /contact-us/.

Last updated: 1 November 2026.

References

1. Deprecations — OpenAI API

1. Model deprecations — Claude Platform Docs

1. Gemini deprecations, last updated 22 October 2026 — Google AI for Developers

1. Model versions and lifecycle, last updated 30 October 2026 — Google Cloud

1. Model lifecycle — Amazon Bedrock User Guide

1. Foundry Models lifecycle and support policy, last updated 24 July 2026 — Microsoft Learn

1. Model lifecycle (Legacy) — Amazon Bedrock User Guide

1. Model retirement schedule, last updated 23 September 2026 — Microsoft Learn

1. List models, API reference — Claude Platform Docs

1. Models overview — Claude Platform Docs

1. Migrating to API version 2026-10-21 — Claude Platform Docs

1. Claude Platform release notes, entry of 21 October 2026 — Anthropic

1. Gemini models, last updated 28 October 2026 — Google AI for Developers

1. Gemini API release notes, entry of 8 October 2026 — Google AI for Developers

1. Partner model deprecations, last updated 30 October 2026 — Google Cloud

1. Anthropic Claude Haiku 5.5 model card — Amazon Bedrock User Guide

1. OpenAI GPT-6.1 Sol model card — Amazon Bedrock User Guide

1. Anthropic Claude Opus 4.6 model card — Amazon Bedrock User Guide

1. Anthropic Claude Opus 5.5 model card — Amazon Bedrock User Guide

1. Anthropic Claude Sonnet 5.5 model card — Amazon Bedrock User Guide