Why the notice reaches the wrong person
A deprecation notice is addressed to an account, not to a codebase. OpenAI says it notifies customers who are actively using the model by email and documents the change on its deprecations page.1 Microsoft sends the GA retirement notice at least 60 days before retirement, to subscription owners with active deployments.2 In most companies the person who owns the account or subscription is not the engineer who wrote the call, and not the team that still runs a nightly job against gpt-4o-mini.
The fix is not a better inbox rule. The vendors already publish the same facts in places a script can read, and your side of the join (which model ids you send, and who owns them) is data you hold. A scheduled check does the join every day and turns each hit into a ticket with an owner and a date.
This is the piece the usual artifacts do not automate. A deprecation calendar is a human-maintained list, and a migration playbook says what to do once you know. Neither tells you that a service you forgot about calls an id that appeared on a vendor page this morning.
What the check has to do
Keep the job small and make each stage testable on its own.
- Inventory: list every model id your applications call, with an owner and a source (repo, config, gateway key).
- Collect: pull each vendor's lifecycle data, preferring APIs over pages.
- Normalize: reduce ids to a comparable form.
- Join: match inventory rows to vendor rows, per host.
- Score: turn the shutdown date into a severity.
- File: open or update one ticket per owner and model, and stay quiet when nothing changed.
The ordering matters because failures are different at each stage. A bad inventory gives you silence about a model you really use. A bad parser gives you silence about everything. A bad join gives you false alarms, which teach people to ignore the channel. The sections below take the stages in that order of risk.
Build the inventory from two sides
Static scanning and traffic data each miss things the other catches, so use both.
Static: grep repositories, configuration, infrastructure code and prompt files for model id strings. This finds code paths that have not run recently, such as a quarterly report job. It misses ids assembled at runtime or read from a database.
Observed: read the model field from your request logs for the last 30 days. This finds the dynamic cases and the forgotten service that still runs. It misses anything that has not sent a request in the window. If your traffic goes through a gateway, an export of the request log gives you model, key and project per request; INFRO's request log export and per-request usage record exactly those fields.
Write both into one file with one row per (model id, caller):
{"model": "gpt-4o-2024-08-06", "host": "openai", "owner": "team-search", "source": "gateway-key:search-prod"}
{"model": "gpt-4o", "host": "azure-openai", "owner": "team-support", "source": "repo:support-bot/config.yaml"}Record the host next to the id. The same model name retires on different dates depending on where you call it, and the join in a later section depends on that. Record unpinned aliases as they appear (gpt-4o, not a guess at a snapshot), because an alias is itself a finding; see aliases versus pinned snapshots.
Collect from APIs first, pages second
Prefer a field that returns a date over a table you have to parse.
OpenAI. The deprecations page recommends monitoring programmatically by calling GET /v1/models and checking each model's optional shutdown_date field, which is also on the retrieve-a-model endpoint.1 Because the field is optional, its absence means "no date published on this endpoint", not "safe". Also read the page itself: it carries a table of shutdown date, affected model, status, recommended replacement and notice date, a separate table for feature and endpoint deprecations, and a note that fine-tuned versions follow the schedule of their base model.1 The page has a markdown version at the same address with .md appended, which is easier to parse than the rendered HTML.
Azure OpenAI and Foundry. Microsoft says you can check lifecycleStatus, deprecation and per-SKU deprecationDate for any model through the Models API, scoped to a subscription and a region.2 The REST reference lists a List operation for it.3 The retirement schedule page is the human-readable twin, with lifecycle, retirement date and replacement per model and version.4 Use the API as the source of truth and the page as a cross-check.
Other vendors. Add each one the same way: look for a list endpoint with a date field, fall back to a page. For a vendor whose page you cannot parse reliably, store a hash of the deprecation table and open a review task when it changes.
import re, datetime as dt, urllib.request
DATE_ROW = re.compile(r"^\|\s*\*\*(?P<date>[A-Z][a-z]{2} \d{1,2}, \d{4})\*\*\s*\|(?P<ids>[^|]*)\|", re.M)
TICKS = re.compile(r"`([^`]+)`")
def openai_page_rows(md: str):
rows = []
for m in DATE_ROW.finditer(md):
shutdown = dt.datetime.strptime(m["date"], "%b %d, %Y").date()
for mid in TICKS.findall(m["ids"]):
rows.append({"host": "openai", "id": mid, "shutdown": shutdown})
return rows
def fetch(url, headers=None):
req = urllib.request.Request(url, headers=headers or {})
with urllib.request.urlopen(req, timeout=30) as r:
return r.read().decode()
md = fetch("https://developers.openai.com/api/docs/deprecations.md")
rows = openai_page_rows(md)
# Fail loudly: a layout change must not look like "nothing is retiring".
assert len(rows) >= 10, f"parser returned {len(rows)} rows; page layout changed?"The assertion at the end is the most important line. A table that moves, gains a column or becomes a script-rendered widget will make the regex match nothing, and a job that reports "0 hits" will look healthy.
Normalize, then join per host
Vendors write ids in several shapes: a family label with snapshots in parentheses, a dated snapshot, a bare alias. OpenAI's table, for example, lists GPT-4o as a family label with the snapshot ids gpt-4o-2024-08-06 and gpt-4o-2024-11-20 beside it.1 Azure lists a model name and a version in separate columns.4 Reduce everything to a (base, snapshot) pair and match on the base when your inventory holds an alias.
import re, datetime as dt
SNAP = re.compile(r"-(\d{4}-\d{2}-\d{2})$")
def split_id(model_id: str):
name = model_id.lower().split("/")[-1]
m = SNAP.search(name)
return (SNAP.sub("", name), m.group(1) if m else None)
def match(inv, vendor_rows):
base, snap = split_id(inv["model"])
hits = []
for r in vendor_rows:
if r["host"] != inv["host"]:
continue # never compare across hosts
rbase, rsnap = split_id(r["id"])
if rbase != base:
continue
if snap is None or rsnap is None or snap == rsnap:
hits.append(r)
# An unpinned alias is exposed to the earliest matching shutdown.
return min(hits, key=lambda r: r["shutdown"]) if hits else NoneTwo rules in that function come from the vendors' own data.
Compare per host. The same model can retire on different dates on different providers. OpenAI lists gpt-4o-2024-08-06 and gpt-4o-2024-11-20 shutting down on Jan 14, 2027,1 while Azure's schedule has those versions retiring on April 14, 2027, and gpt-4o version 2024-05-13 on January 14, 2027.4 A join that ignores host reports the wrong date for one of your two callers. Azure also distinguishes deployment types: Global Standard and Data Zone Standard deployments are auto-upgraded at retirement while provisioned deployments are not.2 Carry the deployment type in the inventory if you use more than one.
Treat aliases as the earliest date. If an app sends gpt-4o with no snapshot, it is exposed to whichever snapshot goes first.
Beware of Azure deployment names. A deployment is a name you chose, and it may not equal the model name. Resolve each deployment to its model and version through the Azure control plane before the join, or the inventory will list a name no vendor row will ever match.
Turn a hit into an owner, a date and a severity
A match is not yet actionable. The ticket needs three things.
Owner. Resolve in this order: the gateway key or project owner for observed traffic, then CODEOWNERS for the file where the static scan found the id, then the service catalog entry. If nothing resolves, route the ticket to the platform team's own queue so that the unowned case is visible, not dropped.
Date. Use the shutdown date from the vendor row, and keep the status and replacement in the ticket body. Notice windows differ, which is why one threshold does not fit every id. OpenAI states minimum notice of at least 6 months for generally available models, at least 3 months for specialized variants and image generation models, and possibly as little as 2 weeks for preview models.1 Microsoft gives at least 60 days for GA retirement and at least 30 days for preview.2
Severity. Base it on days remaining and on whether the id is a preview:
def severity(days_left: int, preview: bool) -> str:
if days_left <= 0:
return "shut-down" # already failing or refused
if preview:
return "critical" if days_left <= 30 else "warning"
if days_left <= 14:
return "critical"
if days_left <= 60:
return "warning"
return "info"Make filing idempotent. Key each ticket on (host, base id, shutdown date) and update it when the date or replacement changes. A re-run must neither duplicate tickets nor stay silent when a vendor moves a date earlier. Re-notify when severity steps up, not on every run.
Failure modes to design against
- Silent parser breakage. Covered above: assert a minimum row count and alert when it drops. Also alert when a vendor's source is unreachable for several consecutive runs.
- Dates that move. Microsoft notes the schedule is subject to change and that its replacement is chosen roughly 90–120 days before retirement,2 so a ticket created early may name no replacement yet. Update rather than close.
- Unversioned ids and family labels. A hit on a family label with no snapshot match is still a hit; file it for review.
- Fine-tunes. OpenAI says fine-tuned versions follow their base model's schedule.1 Microsoft retires fine-tuned training and deployment in separate phases.4 Add fine-tune ids to the inventory explicitly; they will not look like the base name.
- Non-model deprecations. Endpoints and features retire too: OpenAI's page lists prompt objects, Agent Builder, audio translations and fine-tuning job creation on their own dates.1 Inventory the features you call, not only the models.
- Alert fatigue. Notify on change and on severity step-ups only. A daily digest of everything still open belongs in the weekly review, not the on-call channel.
What a gateway already covers, and what it does not
If your calls pass through INFRO, part of this is already recorded for you. INFRO records each vendor's deprecations, shutdown dates and recommended replacements from the vendor's own page; GET /v1/models carries a retirement object, and an organization that sent a retiring model in the last 30 days gets a model.retiring alert naming the keys and projects that called it.5 The alert fires when a model the organization sent is shut down by its vendor within 60 days, and is critical in the last 7 days.6 INFRO also reads the model list your own key returns every three hours and stops offering a model the key has stopped listing, and the console shows these notices under the bell. See the models and lifecycle docs (opens in a new tab) for the exact fields.
That covers traffic that goes through the gateway. It does not read your repositories, calls that bypass the gateway, or your Azure deployment names, and INFRO records a vendor's retirement for a partner cloud only when that cloud's own date is recorded.5 So the scheduled job in this guide still earns its place: it covers the static inventory and the owner mapping, and it can ingest the gateway's retirement data as one more source instead of scraping the same pages twice.
A rollout checklist
- Static scan of repos and config produces
(model, host, owner, source)rows. - Request-log export produces the same rows from the last 30 days.
- Azure deployment names are resolved to model and version.
- OpenAI
GET /v1/modelsshutdown_dateand the deprecations page are both read. - Azure Models API lifecycle fields are read and cross-checked with the schedule page.
- Each collector asserts a minimum row count and alerts when a source is unreachable.
- Join is per host; unpinned aliases map to the earliest matching date.
- Tickets are keyed on
(host, base id, shutdown date)and updated, not duplicated. - Unowned hits go to the platform queue, never to nobody.
- Each ticket links to your migration playbook and the calendar entry.
