Refolk
ReferenceEngineering and open source

The Vendor Retirement-Signal Reference: Proven Date, Scope, and Trust

Look up any vendor retirement signal and write down the date it proves, its scope, whether the vendor can move it, the confirming check, and a confidence grade.

22 min readLast reviewed November 20, 2026Read as Markdown

Key takeaways

  • Under RFC 9745 a Deprecation header is only a hint that changes no behavior, and its value must be an @-prefixed Unix timestamp. GitHub has been captured sending an HTTP-date instead, which a strict parser drops without any warning.
  • Published preview floors are far shorter than GA floors: OpenAI says preview models may get notice "such as 2 weeks", and Microsoft Foundry promises at least 30 days for previews against 60 for GA.
  • Stated notice floors tend to become the schedule: Anthropic gave 61 days for Sonnet 4.5 and 60 days for Haiku 3 against a 60-day floor, and OpenAI's GPT-4o and GPT-4.1 notice was reported at 92 days against a six-month floor.
  • In endoflife.date the support field is inverted: support: false means the release is no longer supported and the exact end date is not known.
  • The EU Cyber Resilience Act support-period end date is the only legally binding retirement date a consuming team will see. It must run at least five years, and its month and year must be stated at purchase.
  • In Refolk's index there are 19 US profiles titled API Product Manager against 284 US developer-relations profiles. The people who own deprecation policy are about 15 times scarcer than the people who announce it.

This reference is for engineering managers, platform leads, technical founders, and developer-relations leads who run a vendor API, SDK, model, or product version in production and need to know the real date it stops working. Each entry covers one retirement signal a vendor can publish. It records the date the signal proves, what that date covers, whether the vendor can move it, how the signal misleads, and the check that confirms it. The last section turns those entries into a weekly procedure that ends with a graded migration calendar.

Which retirement signals bind the vendor, and which only hint?

Almost none of them bind. Of the signals a consuming team sees, only a support-period end date declared under the EU Cyber Resilience Act (CRA) has legal force. Every voluntary date is a hint, an expectation, or a promise the vendor can soften later.

Use this table as the index. Each row has a fuller entry further down.

SignalDate it provesVendor can move it?Confirming check
Deprecation headerDeprecated as of (a hint)YesLog the raw header string
Sunset headerExpected shutdownYesMatch the lifecycle page
OpenAPI deprecated: trueNo date at alln/aFind a header or page row
Vendor lifecycle page rowTier floor; set or tentative dateYes, via carve-outsModel ID and tier label
endoflife.date rowCommunity copy of a dateYes, by pull requestVendor page wins
CRA support periodMonth and year, 5+ yearsNoPurchase-time declaration

The wording matters more than the date, so read every date through the verb attached to it:

  • RFC 9745 calls the Deprecation header "a hint about deprecation to the resource consumer".
  • RFC 8594 describes the Sunset date as the date the provider expects the resource to stop responding.
  • OpenAI's notice floors apply "unless safety or compliance concerns require a faster timeline".
  • Google's 2012 policy rewrite turned a commitment into "we will strive to provide one year notice".

How much each retirement signal binds

  1. Law
    CRA support-period end date, month and year declared at purchase
  2. Vendor's set date
    Retirement date on the vendor's own lifecycle page, not labelled tentative
  3. Vendor notice policy
    Per-tier minimum notice, with carve-outs and scope wording
  4. Protocol hints
    Deprecation and Sunset headers on a single resource
  5. Community data
    endoflife.date rows, edited by pull request
The lower a signal sits in this stack, the less it binds the vendor and the more checking it needs before it goes on a calendar.

Deprecation header vs Sunset header: what the HTTP signals mean

The Deprecation header says a resource is or will be deprecated, and it changes nothing about how the resource behaves. The Sunset header says when the provider expects the resource to stop responding. Both apply only to the resource that returned them, and neither commits the provider to anything.

Deprecation header (RFC 9745)

  • Status: RFC 9745 is a Proposed Standard, published March 2025.
  • Format: the value must be a Date in the form @1688169599, which is an @ followed by a Unix timestamp.
  • Date it proves: when the resource was or will be deprecated. The value may be in the past or the future.
  • Scope: only the resource that returned it. It says nothing about the rest of the API or the product.
  • Can the vendor move it: yes. It is a hint, and per the RFC "the act of deprecation does not change any behavior of the resource".
  • How it lies: earlier drafts of the spec allowed the value true and allowed an HTTP-date, and code written against those drafts still sends them. One practitioner capture shows GitHub sending Deprecation as an HTTP-date. A strict RFC 9745 parser drops that value, so you see no signal at all.
  • Check: log the raw header string next to whatever your parser produces.

Sunset header (RFC 8594)

  • Status: RFC 8594 is Informational, published May 2019. Informational means it is published for reference rather than as a standard.
  • Format: an HTTP-date, the same format as the standard Date header. It should be in the future.
  • Date it proves: when the provider expects the resource to stop responding. "Expected" is the strongest word it offers.
  • Scope: only the resource that returned it. RFC 8594 also defines a rel="sunset" link for scope wider than one resource.
  • Ordering rule: RFC 9745 says the Sunset timestamp must not be earlier than the Deprecation timestamp. A pair that breaks this rule is malformed. Flag it instead of picking one of the two dates.
  • Lead time: no public figure for typical lead time exists.
  • Check: match the date to a row on the vendor's lifecycle page.

OpenAPI deprecated: true

  • Proves: the operation or schema is marked deprecated in the spec. The flag carries no sunset date. The header does.
  • How it lies: teams read it as "going away soon" or wave it off as "someday". The flag supports neither reading.
  • Check: look for a Sunset header on the live response, or a row on the lifecycle page.

Field-level changes get no HTTP signal at all. RFC 9745's header covers whole resources, so when a field changes shape, the changelog is the only place the change appears.

410 Gone

  • Proves: the resource has already stopped working. It is the end state, not a notice.
  • Example: Microsoft Foundry gives previews that have no replacement 30 days' notice, after which inference returns 410 Gone.
  • Use: treat any 410 from a vendor host as an incident. Then work backwards to find which earlier signal you missed.

SDK warnings, brownouts, archived SDK repositories

A brownout is a temporary, scheduled shutdown that a provider runs before the final date so that callers notice. No public source establishes typical lead times for brownouts, SDK runtime deprecation warnings, or archived SDK repositories. Treat each one as grade D evidence that a date exists somewhere, and then find that date on the lifecycle page.

Vendor lifecycle pages and published notice floors

A vendor's lifecycle page states the minimum notice it promises for each tier. Row by row, it also lists the deprecation and retirement dates it has set. The floor depends heavily on tier: a preview can get as little as two weeks' notice, while a generally available (GA) model gets months.

ProviderGA floorPreview floor
OpenAI6 months (3 for specialized variants)"such as 2 weeks"
Anthropic60 daysNot stated separately
Microsoft Foundry60 days30 days
Vertex AI, short-lived Gemini45 daysn/a

The OpenAI figures come from a secondary capture of OpenAI's own deprecations page, and the Vertex figure comes only from secondary sources. The Anthropic and Microsoft Foundry figures come from the vendors' own pages. Before filing any of these, re-read the live page, because the floor is policy text the vendor can rewrite.

Reading one row

  • Tier: GA, specialized variant, or preview. OpenAI ties the preview tier to model names that contain "preview", so read the model ID string itself.
  • Scope wording: Anthropic's 60-day floor covers "publicly released models". Restricted models fall outside it.
  • Carve-outs: OpenAI can shorten its floor for safety or compliance reasons, so treat any OpenAI date as movable.
  • Date status: Anthropic's Claude Sonnet 4.5 retirement date was labelled "tentative". claude-mythos-preview was deprecated with its retirement "To be announced". A deprecation with no retirement date only proves that the clock has not started yet.
  • Forced upgrades: Microsoft Foundry can force-upgrade preview deployments with at least 30 days' notice. Your behavior changes even though nothing retires.
  • Platform: Bedrock and Vertex retire models on their own schedules. According to a secondary source, Bedrock uses "a Legacy period of 6 months or 45 days" for newer models. A model vendor's date does not carry over to a reseller cloud.

Changelog posts, emails, and contract clauses

No public source establishes typical lead times for changelog posts or deprecation emails. Treat each one as a pointer: wait to file the date until you have matched it to the lifecycle page, because the page is where the tier and the floor are stated. If your contract names a notice period or support term, file it as its own row and quote the clause word for word. There is no public evidence on how such clauses are usually worded, so grade it by its verbs, the same way you would a policy page.

Grading a date by wording and commitment

Published floor or lawHint or "strive"
Floor exists, date not set
Grade C; plan to today plus the tier floor
Set date backed by floor or law
Grade B, or A if law; file the date
Hint with no settled date
Grade D; plan to the shortest floor and escalate
Set date backed only by a hint
Grade C; confirm on the lifecycle page
Tentative or to be announcedSet retirement date
A date is only as strong as the weaker of two things: how settled its wording is, and what commitment stands behind it.

How much notice before API shutdown, in practice?

Expect the floor, and sometimes less. Vendors that publish a minimum tend to give exactly that minimum. In at least one reported case the notice was about half the stated floor.

CaseStated floorGivenMargin
Anthropic Sonnet 4.560 days61 days+1 day
Anthropic Haiku 360 days60 days0
OpenAI GPT-4o / 4.1~182 days92 days-90 days (~51%)
Google App Engine policy3 years1 year-67%

I calculated the margins myself. The OpenAI case is reported only by secondary sources, which give a 92-day window from October 14 to January 14, set against a six-month minimum for GA models.

92 days
Reported notice for GPT-4o and GPT-4.1 retirement, against a six-month GA floor
Secondary sources only, but enough to show that a published floor is not a guarantee, so plan as though any date could arrive early.

Short notice hurts because teams find out slowly. One secondary source reports that in 94% of applications studied, the model name was written directly into the code. The same source reports that 89% of applications were fixed late on Anthropic notices of 60 to 114 days, against 13% on OpenAI's one-year Assistants notice. The original study is not cited. Treat these numbers as a direction, not a rate you can plan with. The direction is clear, though: when notice is short and model IDs are hardcoded, migrations land late.

Policies can also shrink without any date changing. On April 20, 2012, Google announced that App Engine, the Maps/Earth APIs, and the YouTube API would move from a three-year deprecation policy to a one-year policy in April 2014. App Engine called one year "the longest deprecation period Google offers across APIs and developer products". For other APIs, Google wrote "we are removing the deprecation policy - not the API". No single endpoint's date moved, so a monitor watching endpoints would have seen nothing.

A published minimum becomes the schedule. Plan for the floor, and for the day the floor is rewritten.

Who at the vendor owns a retirement date?

Usually an API product manager owns the date, not the developer advocate who announces it, and that is a small group of people. In Refolk's index of professional profiles, US developer-relations profiles outnumber US API Product Managers by about 15 to 1.

Segment (Refolk's index)ProfilesTop employer share
API Product Manager, US19U.S. Bank, 7 of 19 (37%)
API Product Manager, UK5Swift, 2 of 5 (40%)
Developer Relations, US284Google, 3 of a 25-profile sample
15x
US developer-relations profiles per US API Product Manager in Refolk's index
284 against 19; the people who set deprecation policy are a much smaller group than the people who announce it.

This matters when a row is graded C or D: a tentative date, a retirement marked "To be announced", or a header that disagrees with the page. Developer advocates can pass a question along, but the policy owner can actually answer it. The counts come from exact title matching, so the real population is likely larger. Variant titles will not show up in these numbers. US API Product Managers outnumber UK ones 3.8 to 1, and both groups are concentrated at a few employers, which makes them findable once you know the title to search for.

Searching for that title by hand across the public web is slow. Refolk takes the request in plain English and returns named people with their current employer, so you can go to the policy owner directly instead of filing another support ticket.

Date-confirmation note to a vendor's API product owner
Subject: Retirement date for <resource or model ID>

Hi <name>,

My team runs <resource or model ID> in production, called through <direct API or platform name>.

What I can see today:
- Lifecycle page wording: "<exact wording, e.g. tentative or To be announced>", captured <date>
- Raw headers: Deprecation <value>, Sunset <value>
- Tier as I read it: <GA / specialized variant / preview>

Two questions:
1. Is <date> a set retirement date, or could it move earlier?
2. Does this tier's notice floor apply on the platform we call it through?

A one-line answer is plenty. Thank you.

Replace the angle-bracket fields, and send it only after you have checked the lifecycle page and the raw headers.

End of support vs end of life vs end of sale in endoflife.date

In endoflife.date, support marks the end of active support, eol marks the point where all support stops (security support included), and there is no end-of-sale field. discontinued is the nearest match to end of sale. Vendors define these terms on their own pages, and this reference did not collect their definitions. Read the vendor's own wording before you map it onto these fields.

endoflife.date is a community-maintained dataset of product release cycles. Each product row can carry these fields: releaseCycle, releaseLabel, codename, releaseDate, lts, support, eol, extendedSupport, discontinued, latest, latestReleaseDate, and link.

FieldAPI v1 nameWhat it means
supportisEoas, eoasFromActive support ends; false means ended, date unknown
eolisEol, eolFromAll support stops, including security
discontinuedNot covered hereNearest thing to end of sale
  • Proves: a contributor copied a date from somewhere. Contributors edit Markdown front matter by pull request, under guiding principles published on the project's wiki. See the contributing guide.
  • Can it move: yes, whenever someone merges a pull request.
  • How it lies: support: false reads like "not yet ended" but means the release is no longer supported and the exact date is unknown. A date in the dataset can also disagree with the vendor's own page.
  • Check: the vendor page is the source of truth. Log every disagreement, and file the vendor's date.

The project's size and how often its dates disagree with vendor pages are not established publicly. If you depend on it, measure your own disagreement rate across the products in your inventory.

The CRA vendor support period end date

Under Article 13 of the EU Cyber Resilience Act, a manufacturer must set a support period of at least five years, or the product's expected use time if that is shorter, and state the end month and year at the time of purchase. It is the only retirement date in this reference with legal force behind it.

  • Article 13(8): "the support period shall be at least five years". If the product is expected to be in use for less than five years, the support period matches the expected use time.
  • Article 13(19): the end date of the support period, "including at least the month and the year", must be clearly specified at the time of purchase.
  • Article 13(9): updates stay available for at least ten years after release, or for the rest of the support period, whichever is longer.
  • When it applies: according to a secondary source, these obligations apply from 11 December 2027.
  • Versions: guidance reportedly requires each substantially modified version to have its own declared support period. Record the date per version, not per product line.

No published method exists for recording a CRA date next to a voluntary lifecycle date. Keep them as two separate rows and never merge them. If a vendor's voluntary end-of-life date falls before its declared CRA support-period end for an in-scope product, take that to the vendor and to counsel. Do not settle it on the calendar yourself. Whether a given product is in scope is a legal question, so get that answer before you grade anything A. See the Article 13 text.

How retirement dates mislead: failure modes and false positives

Most missed retirements come from misreading a signal you already had: a format your parser dropped, a scope read too wide or too narrow, or a soft date filed as firm. Each failure below names what happens and the check that catches it.

  1. Deprecation header sent as an HTTP-date. A strict RFC 9745 parser drops it, and the dashboard shows nothing. Check: log the raw header string, not only the parsed value.
  2. Scope read too wide or too narrow. A header on one endpoint gets read as covering the whole API. Going the other way, a field-level change arrives with no header at all. Check: look for a rel="sunset" link or a changelog entry before widening or narrowing scope.
  3. A "tentative" or "not sooner than" date filed as firm. The calendar shows a hard date the vendor never committed to. Check: read the exact wording on the lifecycle page, and keep the grade at C until a set retirement date appears.
  4. Preview tier mistaken for GA. You plan for months of notice and get weeks. OpenAI ties the preview tier to model names containing "preview". Check: read the model ID string, not the marketing name.
  5. Reseller clouds on their own schedule. The model vendor's date gets applied to Bedrock, Vertex, or Foundry deployments. Check: look up each platform's lifecycle page separately, and file one row per platform.
  6. support: false read as "still supported". The endoflife.date boolean is inverted. Check: confirm against the vendor page before acting on any dataset value.
  7. Policy softened from "will" to "strive". The floor disappears without any date changing. Google's 2012 rewrite is the documented example. Check: diff dated captures of the policy page.
  8. Safety or compliance carve-out. OpenAI's floor can be shortened for these reasons. Check: nothing confirms an OpenAI date is safe from this. Treat it as movable and start the replacement evaluation when notice arrives.

Two compounding factors make every one of these worse. The first is hardcoded model IDs: per the secondary study above, 94% of applications studied wrote the model name directly into the code, so finding where a model is used takes longer than the notice period allows. The second is policy text that no one watches, because an endpoint monitor cannot see a policy page being rewritten.

How to track vendor API deprecation dates, step by step

Build an inventory, capture signals at three layers (response headers, vendor pages, and policy text), grade each date, and file it. No public source documents this procedure end to end. I assembled it from the RFCs, the vendor pages, and the failure cases above.

From vendor signal to graded calendar entry

  1. Inventory every vendor resource
    List each endpoint, model ID, and SDK version in production, including model names hardcoded in source. Done when every item has a named owner, which usually takes about a day.
  2. Log Deprecation and Sunset headers raw
    Have the platform team log the raw strings on every third-party response, accepting both the @-timestamp and HTTP-date formats. Done when the first appearance of either header on any resource raises an alert.
  3. Read the vendor lifecycle page
    Record each item's tier, the stated floor for that tier, and whether the date is set, tentative, or to be announced, with reseller platforms as separate rows. Done when each row has a tier and a floor.
  4. Cross-check endoflife.date
    Compare the dataset row with the vendor page, remembering that support: false means support has ended. The vendor page wins, and the step is done when every disagreement is logged.
  5. Record the CRA support-period date
    For products in scope of the Cyber Resilience Act, record the declared end month and year as its own row. Done when every in-scope product has one.
  6. Snapshot and diff the policy page
    Capture each vendor's notice-policy page with the Wayback Machine or your own capture tool and diff it against the last snapshot, monthly as a starting cadence. Done when any change in floor, tier, or verb has been logged.
  7. Grade and file on the migration calendar
    Assign a grade from A to D, set the planning date by the grade rule, and file the entry. Done when every entry has a grade, a planning date, and a next re-check date.

The grade and the planning date

  • A, binding: a CRA support-period end date, declared at purchase, for a product counsel has confirmed is in scope. Plan to that date.
  • B, set: a retirement date on the vendor's own lifecycle page, not labelled tentative, checked on the platform you actually call. Plan to that date, and note any carve-out.
  • C, expected: a Sunset header, a "tentative" or "not sooner than" date, or a date seen only in a changelog post or email. Plan to whichever comes first: the stated date, or today plus the tier floor.
  • D, hint: a Deprecation header alone, an OpenAPI flag, an endoflife.date value with no vendor confirmation, or a retirement "To be announced". Plan to today plus the shortest floor that could apply, and escalate.

The sources disagree on when to act. One header-focused guide says to act as soon as the header appears. The guides on AI model retirement say to start evaluating the replacement as soon as notice arrives. I do both: the header triggers the alert, and any notice graded C or better triggers the replacement evaluation that same week, because the evaluation is what takes time.

Migration calendar entry
Resource:
Platform called through:
Owner:
Signal type: Deprecation header / Sunset header / lifecycle row / changelog or email / endoflife.date / CRA / contract clause
Raw signal (verbatim):
Date it proves:
What the date covers:
Tier and stated floor:
Can the vendor move it, and why:
Confirming check done (what, when):
Grade: A / B / C / D
Planning date:
Replacement eval started: yes / no
Next re-check:

One entry per resource per platform. Paste the raw signal text, never a paraphrase.

Before calling the calendar current

  • Every vendor resource in production has a named owner.
  • Raw Deprecation and Sunset strings are logged on every third-party response, in both formats.
  • Each lifecycle row records tier, stated floor, and date status: set, tentative, or to be announced.
  • Bedrock, Vertex, and Foundry deployments each have their own row, separate from the model vendor's.
  • Every endoflife.date value has been checked against the vendor page, with disagreements logged.
  • Every in-scope product has its CRA support-period end month and year on a separate row.
  • Each vendor's notice-policy page has a dated snapshot from this month, diffed against the last.
  • Every calendar entry carries a grade, a planning date, and a next re-check date.
  • No grade C or D entry has a planning date later than today plus its tier floor.

Keeping the calendar current

Re-run steps 2 to 7 weekly for anything graded C or D, and monthly for everything else. On every pass, re-grade any row whose wording has changed. A move from "tentative" to a set date raises the grade, and a move from "will" to "strive" in the policy text lowers every row that depends on that policy. If a grade D row has gone two re-checks without getting a date, ask Refolk for the vendor's API product managers and send them the confirmation note above. The monthly policy-diff cadence is a starting point, not a tested figure. Tighten it for any vendor whose policy page has changed between two consecutive snapshots.

Questions practitioners ask

What is the difference between the Deprecation and Sunset headers?

The Deprecation header (RFC 9745) says a resource is, or will be, deprecated, and deprecation changes nothing about how the resource behaves. The Sunset header (RFC 8594) gives the date the provider expects the resource to stop responding. Deprecation takes an @-prefixed Unix timestamp and Sunset takes an HTTP-date. The Sunset date must not be earlier than the Deprecation date. Both headers cover only the resource that returned them, and neither one binds the provider.

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

Published floors range from months for generally available models down to weeks for previews. OpenAI states at least six months for GA models, three months for specialized variants, and "such as 2 weeks" for previews. Anthropic and Microsoft Foundry each state 60 days for GA, and Foundry states 30 days for previews. In practice notice tends to match the floor exactly. One reported OpenAI case gave 92 days against a six-month floor.

Can I rely on endoflife.date for vendor end-of-life dates?

Use it as a cross-check, not as the source of truth. endoflife.date is maintained by its community through pull requests, and its dates can disagree with the vendor's own page. Its support field is also inverted: false means the release is no longer supported and the exact date is not known. When the dataset and the vendor page disagree, file the vendor's date and log the disagreement.

What is the difference between end of support, end of life, and end of sale?

In endoflife.date's schema, the support field marks the end of active support and eol marks the point where all support stops, security support included. There is no end-of-sale field, and discontinued is the closest match. Each vendor defines these terms on its own pages, and the definitions vary. Read the vendor's own wording before you map its terms onto your calendar.

Can a vendor shorten its deprecation notice policy?

Yes, so watch the policy itself as well as the dates. On April 20, 2012, Google announced that App Engine, the Maps/Earth APIs, and the YouTube API would move from a three-year deprecation policy to a one-year policy in April 2014. The new wording was "we will strive", and Google dropped the policy entirely for other APIs. No endpoint's date changed, so only a diff of the policy page would have caught it.

How do I monitor third-party API sunset dates automatically?

Log the raw Deprecation and Sunset header strings on every third-party response. Accept both the @-timestamp and HTTP-date formats, and alert the first time either header appears for a resource. That only catches notices at the resource level. Field-level changes send no header, reseller platforms keep their own schedules, and policy cuts never show up in a response. Pair header logging with lifecycle-page checks and monthly policy-page snapshots.

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next