Guide · model lifecycle

AI model deprecation policy and migration

Every provider retires models on a different clock. This page compares the published notice periods, lists dates already on the calendar, and gives a migration checklist that survives an alias change.

Short answer. Pin a versioned model for production, monitor the provider's deprecation page and your own model inventory, and keep a tested replacement ready. Notice periods range from about two weeks to twelve months. A “latest” alias is convenient, but it can change behaviour without a deploy.

What a deprecation policy must answer

A useful policy answers four questions: how much notice you get, how you are notified, what happens to requests after the date, and whether you can keep the old model through a dedicated deployment. Most provider pages answer the first two and are vague about the last two. Read the model card, not just the marketing page.

QuestionWhy it mattersWhere to look
Notice periodSets the migration runwayDeprecations or lifecycle page
Notification channelEmail can go to a dead aliasEmail, status page, API field
Post-retirement behaviourError, redirect or silent re-pointModel card and changelog
Extension optionDedicated capacity may outlive the dateSales terms or contract

The 12-provider notice table

These are the periods each provider publishes, not the notice every retirement actually received. A provider can shorten a window for safety, security or low usage. Check the source page before planning a release.

ProviderPublished noticeAt retirement
OpenAIGA 6 months; specialized 3 months; preview ~2 weeksRequests fail; shutdown_date in models API
AnthropicAt least 60 days for public modelsRequests fail
Google (Vertex / Gemini API)Vertex 12 months after release or 45 days for short-term; Gemini API no fixed minimum, preview ~2 weeksRequests fail or route
Microsoft Foundry / Azure OpenAI18-month GA lifecycle; 60 days notice; preview 30 days410 Gone; some deployments auto-upgrade
AWS Bedrock12 months + 6 months Legacy before 7 Sep 2026; per-card 6 months or 45 days afterRequests fail; no automatic migration
MistralGA 6 months; Labs, Preview and third-party 1 month404
CohereNo fixed period; existing users keep access until shutdownRequests fail
Together AI2–3 weeks serverless; preview under 24 hours after 30 daysRemoved from serverless
GroqNo fixed period; dated notice per modelRequests fail
Fireworks AIAt least 2 weeks for serverlessRequests fail
DeepSeekNo standing policy; changelog onlyRequests fail or temporarily route
xAINo general minimum; 60 days on the 2026 Imagine retirementRetired slugs may redirect

Two lessons: floors are not schedules, and a redirect is more dangerous than an error because the application keeps working with a different model.

Sunset calendar: published dates to watch

The dates below were published as of 20 November 2026. They move, and a provider can add a date after this page is reviewed. Treat the table as a watchlist, not a contract.

DateProviderModel or familyReplacement
30 Nov 2026AnthropicClaude Sonnet 4.5Claude Sonnet 5.5
11 Dec 2026OpenAIGPT-5 family, o3, o3-proCurrent GPT-6 family
14 Jan 2027OpenAIGPT-4o, GPT-4o-mini, GPT-4.1 snapshotsGPT-6 family
14 Jan 2027Azure OpenAIGPT-4o 2024-05-13GPT-5.1
28 Jan 2027Google Vertexgemini-3.7-flashgemini-3.8-flash
1 Apr 2027OpenAIGPT-5.1, GPT-5.3-codex, GPT-5.4-nanoGPT-6 Sol / Luna
8 Apr 2027AWS BedrockClaude Sonnet 4.5Current Claude on Bedrock
30 Apr 2027MistralMistral Large 3, Medium 3.5Mistral Large 4

Earlier 2026 dates, including OpenAI's 23 October batch and Google's 19 November Vertex retirement, are not repeated. Anthropic's 30 November date is still ahead.

Alias versus pinned model behavior

PatternWhat movesBest forRisk
latest aliasPoints to the newest versionPrototypes, internal toolsBehaviour and price change without a deploy
Major aliasNewest minor within a generationTeams that want fixes, not redesignsStill moves under you
Pinned versionNothing until retirementProduction, regulated workloadsYou own every migration date
Dedicated deploymentYour capacity, your upgradeLong-lived or custom modelsCost and operational burden

Pin in production and float in development. Read GET /v1/models and store the returned model IDs in your deployment manifest; the endpoint is the live source of truth for what is available. A CI check that fails when a pinned ID disappears is cheaper than a 3 a.m. page. See the model catalogue.

# Fail deploys if a pinned model disappears from /v1/models
curl -s https://api.plugsky.com/v1/models \
  -H "Authorization: Bearer $PLUGSKY_API_KEY" \
  | jq -e '.data[] | select(.id == "plugsky-pro")' > /dev/null

# A non-zero exit blocks the deploy

Migration checklist

  1. Inventory. Search code, config, notebooks and cron jobs for every model ID and alias.
  2. Pin. Replace aliases with versioned IDs in production; record the retirement date.
  3. Baseline. Run 100–300 representative prompts and store outputs, token counts and latency.
  4. Evaluate. Run the same set on the replacement; compare quality, refusals, JSON validity and cost.
  5. Canary. Send 5–10% of traffic to the replacement with a rollback switch.
  6. Cut over. Move the rest, keep the old path for one retention window, then remove it.
  7. Record. Update the changelog, the runbook and the contract owner. Tell support what changed.

Do not migrate on the shutdown date. The last week is when the replacement is busiest and the least forgiving. A migration that starts 30 days before the date has room for an eval failure; one that starts the day before does not.

Changelog, RSS and monitoring

Plugsky publishes a changelog and an RSS feed at /feed.xml. The current documentation also references a model.deprecated webhook; the webhook section says platform event webhooks are on the roadmap, so confirm availability before you depend on it. Until then, poll GET /v1/models and watch the changelog.

For enterprise commitments, ask for the notice period, the notification address, the post-retirement behaviour and the dedicated-capacity option in writing. Those are contract terms, not documentation guarantees. See the docs for the current API behaviour and the comparisons for provider-by-provider differences.

FAQ

How much notice do AI providers give before retiring a model?

It ranges from about two weeks to twelve months. OpenAI gives at least six months for generally available models, Anthropic at least 60 days, and Mistral six months for GA models. AWS Bedrock, Google Vertex and Microsoft Foundry publish model-specific windows. Some providers publish no fixed period.

What happens when a model is retired?

Most providers reject requests to the retired model ID. Some, including Together AI and xAI, redirect an old ID to a replacement, which changes behaviour and price without a code change. Check the provider's documented behaviour rather than assuming an error.

Does Plugsky publish a model deprecation notice period?

Plugsky publishes a changelog and RSS feed, and the current docs describe model deprecations there. A separate public notice-period commitment is not stated in the pages reviewed on 20 November 2026. Enterprise notice terms should be verified in the contract.

How do I migrate to a replacement model safely?

Inventory every model ID, pin the current version, run a representative eval against the replacement, compare quality, latency and token counts, canary a small traffic slice, then cut over with a rollback path. Remove the old ID only after the retention window.

Sources

OpenAI deprecations page for notice periods and shutdown_date, checked 20 November 2026: developers.openai.com/api/docs/deprecations. Anthropic model deprecations for the 60-day floor, checked 20 November 2026: platform.claude.com/docs/en/about-claude/model-deprecations. AWS Bedrock model lifecycle for the 12-month, 6-month and 45-day windows, checked 20 November 2026: docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html. Mistral model lifecycle policy for GA and preview notice, checked 20 November 2026: docs.mistral.ai/inference/model-lifecycle.

Reviewed 20 November 2026. Notice periods and dates were checked against the providers' own pages on that date and can change. This page is not legal advice; enterprise notice, dedicated capacity and data-retention terms should be verified in the contract with counsel review.