AI API Deprecations: Shutdown Dates & Migration

Find the announced lifecycle conditions for an exact model or API service before planning a migration. This Actor returns 93 complete decision records from selected official OpenAI, Anthropic, Microsoft Foundry and Amazon Bedrock documentation, with provider, version and region scope, related lifecycle stages, replacement guidance, evidence and unresolved conflicts together.

Snapshot observed: 6 November 2026. Main coverage window: 6 November 2026 through 4 February 2027. The snapshot also includes explicitly marked lower-bound and retired control records needed to interpret the schedules. It is a saved document snapshot, with no live endpoint check, scheduled refresh or update SLA. The name “tracker” refers to structured tracking information in this dated release.

Start with one model

Pass this Actor input from the Console, REST API, SDK or an Apify MCP client:

{
  "model_ids": ["gpt-5.5-cyber-preview"],
  "provider_scopes": ["openai_direct"],
  "limit": 1
}

This snapshot returns one scoped decision containing OpenAI's announced 1 February 2027 model shutdown, its source, announcement date and replacement-selection limitation. The event date has day precision and an unknown timezone. The document was observed on 6 November; the run does not test whether an endpoint accepts requests now.

Start the Actor once for the requested input. If your REST or MCP client returns a Run ID before completion, poll that same Run ID and read its default Dataset when it succeeds. Do not start a duplicate Run merely to retrieve its results. Reuse saved records by record_id and record_revision_sha256 when the same snapshot is sufficient.

For an agent workflow, use the exact model ID and provider route, inspect every returned record's version/region/deployment scope and date semantics, then use the provider evidence to plan the appropriate migration. A named replacement is a provider recommendation, not a tested compatible substitute. Some transitions require an API migration as well as a model-name change.

$0.005 per complete decision ($5 per 1,000 results). This example returns one billable item. Set the run's maximum total charge to at least $0.005 for one result. No-match returns no Dataset items and incurs no Dataset-item charge; it does not prove that a model is safe, supported, active or free of deadlines.

What is covered

Route enum	Records	Recorded source scope and limits
openai_direct	31	Selected upcoming deprecation entries relevant to the fixed window, including four service or training workflows. Later schedules outside the window are not an all-model inventory.
anthropic_direct	4	Sonnet 4.5's announced shutdown and three lower-bound controls. A minimum availability date is not an announced shutdown.
azure_foundry	47	Opened retirement-table and lifecycle-stage entries relevant to the window. Model versions, seller routes and deployment footnotes are retained.
amazon_bedrock	11	Relevant scheduled end-of-life entries in the opened Legacy table, an extended-access transition and a regional retired control. Newer model cards were not exhaustively inventoried.

Gemini's current schedule was not verified and is not a selectable provider route in this release. Sources do not establish complete coverage of every provider, active model, region or deadline. A model missing from the results may be outside the covered dates, source sections or supplied filters. No match is never evidence that continued use is safe. This statement also applies when the Dataset is empty.

Input and selection

Field	Meaning
model_ids	Up to 100 exact model IDs or aliases listed in policy_scope.asset_ids. Matching is case-sensitive after Unicode NFKC normalization and outer whitespace trimming; there is no prefix or fuzzy match. Each value is 1–256 printable characters.
provider_scopes	Up to four entries from the exact route enums above. OpenAI direct API and an Azure deployment are different routes.
record_ids	Up to 100 exact IDs from previous results, useful for retrieving a particular version or regional decision.
query	At most 200 characters and 20 whitespace-separated terms. All terms must occur in the saved title, scope, rules or uncertainty. Unicode-normalized, case-insensitive text search.
limit	Integer 1–100, default 1. Omitted filters return a deterministic catalogue-order sample; they do not imply complete coverage.

Values within an ID list are alternatives. Different filter fields combine with AND. Duplicate IDs or aliases for the same decision produce one item. An omitted array does not filter; an explicit empty array selects nothing. Unknown IDs return no matches. Unsupported fields, invalid provider names and malformed values fail before Dataset output.

One ID may legitimately match several records: Azure gpt-chat-latest has distinct model versions, and Bedrock model scopes can differ by region. Use a sufficient limit, inspect selection.truncated_by_limit, then choose exact record_ids if only one scoped decision is needed. The metadata distinguishes matched records from returned records and reports requested model IDs not matched under all supplied filters. It does not represent provider-wide inventory completeness.

Retrieve all four recorded Azure versions for a saved model alias:

{"model_ids":["gpt-chat-latest"],"provider_scopes":["azure_foundry"],"limit":100}

Retrieve a service workflow by text:

{"provider_scopes":["openai_direct"],"query":"self-serve fine-tuning","limit":1}

Service/workflow records have empty asset_ids; service labels are not model IDs and do not match model_ids. Their events distinguish, for example, stopping new training jobs from shutting down fine-tuned inference.

To retrieve the entire current saved corpus:

{"limit":100}

This returns 93 records, costing $0.465 at the stated rate. A $0.50 run cap accommodates this batch. It retrieves the saved selection described above, not all AI models or a fresh scan.

Output and interpretation

Each Dataset item contains record_id, decision_key, title, policy_scope, complete rule_groups, sources, rights, needs_confirmation and confirmation_reasons. The envelope adds coverage_window, source_coverage, selection metadata, a data version, a canonical record hash and separate snapshot and delivery times.

Read each lifecycle_event within its rule group and exact source scope:

• date_kind and date_semantics distinguish announced dates, earliest bounds and conflicting official dates. A lower-bound control keeps effective_date and announced_shutdown_date null; its earliest date remains separate.
• Conflicting official assertions remain together. The Azure gpt-realtime-mini record preserves both published candidate dates and leaves the resolved effective date unknown. The product does not choose the earlier or later date for you.
• timezone: null means the source did not establish a timezone; no timestamp or midnight deadline is invented.
• Deployment exceptions remain material. For Azure gpt-4o, the extension does not revive already retired Regional Standard deployments, and automatic upgrade behavior does not apply to every deployment type.
• Unknown replacements, migration behavior or source coverage stay unknown. A false needs_confirmation flag means there is no recorded additional confirmation reason; it is not an availability or compatibility guarantee.

All related stages and competing claims for the same provider/asset/version/scope are delivered in one complete decision. Shared aliases and snapshots of the same shutdown are not separate charges. Different versions or regional scopes can be distinct decisions. Cross-provider evidence that affects a decision is included without charging a separate row for the evidence itself.

retrieved_at_jst is delivery time. snapshot_observed_at_jst marks the end of a documented observation window, not an exact per-page fetch time. Source observation precision and the window are retained. real_time_lookup, live_endpoint_state_verified, availability_guarantee and replacement_compatibility_guarantee are false.

Pricing and budget behavior

The price is $0.005 per complete decision, including its related stages, conditions, sources and unknowns. There is no Actor-start fee; platform usage is included. Repeated retrieval is charged again. Empty results produce no Dataset-item charges.

The full selected batch must fit the remaining run allowance before output is emitted. If it does not, the run fails with BATCH_BUDGET_EXCEEDED without emitting a partial paid batch. The Actor uses the standard Dataset item event once for each returned decision; it does not separately charge for aliases, source claims or fields. Use record_id plus record_revision_sha256 to cache unchanged records where appropriate.

Evidence and corrections

Official documents remain freely accessible. This product provides hosted, independently structured access to selected factual relationships and their evidence limits. Source links and locators are included in every decision:

• OpenAI API deprecations
• Anthropic model deprecations
• Microsoft Foundry model retirement schedule
• Amazon Bedrock Legacy lifecycle

No source articles, images, private account data or vendor licences are redistributed or claimed. This product is not affiliated with the providers. It does not call model endpoints, change deployments or execute migrations. For a factual error, open an issue with the record ID and the conflicting official evidence.