What retires, and what that means for the calendar
Today is Friday, 2026-11-06. Microsoft's schedule, last updated 2026-09-23 when read, gives four families two dates.1 Retirement means every inference request returns 410 Gone.2
| Deployed model | Version | Retires | Day | Days left | Listed replacement |
|---|---|---|---|---|---|
| codex-mini | 2025-05-16 | 2026-11-15 | Sunday | 9 | none |
| claude-haiku-4-5 | 1 and 2 | 2026-11-15 | Sunday | 9 | none |
| o1, o1-pro, o3, o3-pro, o3-deep-research | various | 2026-11-19 | Thursday | 13 | gpt-5.6-sol |
| o3-mini, o4-mini | various | 2026-11-19 | Thursday | 13 | gpt-5.6-terra |
| Llama-4-Maverick-17B-128E-Instruct-FP8 | none | 2026-11-19 | Thursday | 13 | none |
Three things follow from the table.
First, Nov 15 is a Sunday. If your change process has a weekend freeze, the real deadline for codex-mini and Haiku is Friday Nov 13, which is seven days away.
Second, three of the four families have no listed replacement. Microsoft says it names the official replacement roughly 90–120 days before retirement.2 For these rows that has passed with a dash, so you cannot read an upgrade target off the page and should expect a hard failure, not a swap. The picture is different for the o-series: the replacement is listed, and Microsoft manages automatic upgrades for Global Standard, Data Zone Standard and Standard deployments.2 A deployment you never touched may keep answering on Nov 19, on a different model. I could not confirm how that interacts with each deployment's upgrade option beyond Microsoft's statement that NoAutoUpgrade means the deployment stops working at retirement.3
Third, there is no extension. Microsoft says retirement dates are not extendable and there is no exception process.3
This post sequences the work. For what each deployment type does on the day, see what an Azure OpenAI deployment does on retirement day; for the full dated list, the one-table overview.
The plan at a glance
Run two tracks in parallel from Monday. The Nov 15 track is shorter and has the earlier deadline, so it gets the first canary slot. The Nov 19 track starts its replay on the same Monday, so it is not waiting on the first track.
| Day | Nov 15 track (codex-mini, Haiku 4.5) | Nov 19 track (o-series, Maverick) |
|---|---|---|
| Fri Nov 6 | Inventory, owners, deployment types | Same inventory pass |
| Sat–Sun Nov 7–8 | Optional: freeze replay set from captured traffic | Same |
| Mon Nov 9 | Create candidate deployments; replay unchanged | Create candidates; replay unchanged |
| Tue Nov 10 | Score, tune prompts and schemas; pick target | Score; check cost and latency |
| Wed Nov 11 | Canary at a small share | Tune; pick target |
| Thu Nov 12 | Ramp to full | Canary at a small share |
| Fri Nov 13 | Done; old ids out of rotation; weekend freeze | Ramp |
| Sun Nov 15 | Verify: zero traffic on old ids | Ramp to full on Mon–Tue |
| Wed Nov 18 | Done; old ids out of rotation | |
| Thu Nov 19 | Verify: zero traffic on old ids |
If you have one platform team and not two, do the Nov 15 track first and put the Nov 19 evaluation into Mon–Tue night shifts rather than delaying it. The Nov 19 evaluation is the one that cannot be started late.
Fri Nov 6: inventory every deployment, not every model
The inventory is per deployment. Microsoft's own migration guidance asks for the deployment types (Standard, Provisioned or Batch) of everything the workload uses, with each source model's retirement date and suggested replacement.3
- List every deployment of codex-mini, claude-haiku-4-5 (note both version 1, hosted on Anthropic infrastructure, and version 2, hosted on Azure),1 the o-series ids and Llama-4-Maverick, with subscription, region and deployment type.
- Record each Azure OpenAI deployment's
versionUpgradeOption. Microsoft lists three values (OnceNewDefaultVersionAvailable,OnceCurrentVersionExpired,NoAutoUpgrade).3 - Mark provisioned deployments. They are not auto-upgraded and you migrate them yourself, either in place or side by side.3
- Mark Batch deployments. They follow the side-by-side path: deploy the new model, resubmit jobs, retire the old deployment.3
- Search code, config, feature flags, notebooks, CI evals and scheduled jobs for the old ids and the deployment names. The deployment name your code sends is not always the model name.
- Pull 30 days of request logs by model to find callers nobody listed. Callers that fire monthly are the ones that break on the day.
- Separate fine-tuned deployments. They retire on a separate schedule with training and deployment dates.2 Fine-tuned o4-mini deployments are listed for 2027-11-19, not this year.1
- Name one owner per row, a team that can approve a change.
Capture is the part teams miss. Microsoft notes production content capture is opt-in and never retroactive, so you cannot evaluate traffic you did not record.3 If prompt logging is off, turn it on today for the affected routes, under your logging policy, so Monday's replay has real inputs.
Mon Nov 9: create candidates and replay unchanged
Create the candidate deployments in every region and deployment type that serves traffic, side by side with the old ones so rollback stays possible. Microsoft's Assess phase says to confirm region, SKU and quota, and that the target can run alongside the source.3
Replay your frozen set on the candidate with the current prompts and parameters unchanged. Microsoft's guidance is to replay unchanged before tuning, so you separate behavior the model changed from changes you introduce, and to freeze inputs, ground truths and success criteria for the whole migration.3
# Illustrative harness: replay a frozen JSONL set against two deployments.
import json, time
from openai import AzureOpenAI
client = AzureOpenAI(azure_endpoint="https://YOUR-RESOURCE.openai.azure.com",
api_key="...", api_version="YOUR-API-VERSION")
def run(deployment, messages):
t0 = time.time()
r = client.chat.completions.create(model=deployment, messages=messages)
u = r.usage
return {
"deployment": deployment,
"latency_s": round(time.time() - t0, 2),
"completion_tokens": u.completion_tokens,
"finish": r.choices[0].finish_reason,
"text": r.choices[0].message.content,
}
with open("frozen_set.jsonl") as f, open("replay_out.jsonl", "w") as out:
for line in f:
case = json.loads(line)
for dep in ("o3-old", "gpt-5-6-terra-new"):
out.write(json.dumps({"id": case["id"], **run(dep, case["messages"])}) + "\n")Record the fields that move: verbosity, reasoning depth, structured-output adherence, tool-call shape, refusals and latency variance all commonly change, per Microsoft.3 Keep finish_reason: answers cut off by a length limit are a common regression after a swap.
Picking the replacement, family by family
Microsoft's listed replacement is its recommended target and, where supported, the automatic-upgrade target; it does not limit your manual choice.1 Where nothing is listed, shortlist from rows that are GA on the same page, and look at their retirement dates before you commit.
codex-mini (Nov 15)
No replacement listed. GA candidates on the schedule include gpt-5.1-codex-mini (retires 2027-05-15), gpt-5.3-codex (2027-08-24) and gpt-5.4-mini (2027-09-21).1 Test at least two. If codex-mini sits behind an agent loop, replay real traces: Microsoft says extra fields, renamed arguments and changed sequencing show up only in real traces, not single-turn tests.3
claude-haiku-4-5 (Nov 15, both versions)
No replacement listed. The Claude rows I read on Azure are claude-sonnet-4-6 (2027-02-10) and claude-sonnet-5 (2027-06-30), and I found no newer Haiku on that page.1 Treat Haiku as a latency-and-cost slot: a larger Claude model may answer better and cost more per call, so project cost from your own token logs rather than list rates. Microsoft warns that cost can shift materially between generations.3 If the candidate is not a Claude model, expect calling-code changes too: Microsoft lists API surface, streaming shape and renamed request fields as part of Adapt.3
o-series (Nov 19)
Listed targets are gpt-5.6-sol for o1, o1-pro, o3, o3-pro and o3-deep-research, and gpt-5.6-terra for o3-mini and o4-mini; both retire 2028-01-11.1 Sol and Terra are different tiers, so test the listed pairing and the other tier for your heavier prompts. Reasoning-effort controls don't map one-to-one across generations, and reasoning tokens can swing unit economics by 2x or more.3 Compare cost per completed task, not cost per call.
Llama-4-Maverick (Nov 19)
No replacement listed. The obvious sibling is a trap: Llama-4-Scout and Llama-3.3-70B-Instruct both retire 2027-01-31, 73 days after Nov 19.1 Open-weight rows with longer runway include gpt-oss-120b (2027-09-15) and DeepSeek-V4-Flash (2028-02-20).1 Hosted open-weight models differ between hosts, so a replay matters more here than for the proprietary families.
Tue Nov 10: score, tune, decide
Score the replay against the rubric you fixed on Monday. Tune only after you have the unchanged baseline, and track every change. Microsoft names four surfaces besides prompts: parameters, tool definitions, output schemas and calling code.3
- Parse-failure rate on structured outputs, old versus new.
- Tool-call validity and argument names on real traces.
- p50 and p95 latency, and time to first token if you stream.
- Length-limit cutoffs and refusals.
- Cost per completed task on your own token counts.
- Quota and rate-limit headroom for the target at your peak. Do not assume the replacement's limits match the old model's.
Write the decision down: target id, deployment type, region list, owner, rollback target. If the Nov 15 track fails its bar on Tuesday evening, pick the second candidate and canary it Wednesday rather than holding the date.
Wed Nov 11 to Fri Nov 13: cut over the Nov 15 models
Send a small share of production traffic to the new deployment Wednesday, ramp Thursday, finish Friday. Order the cutover by blast radius:
- Batch and offline jobs first. Resubmit on the new deployment.
- Internal tools and low-volume, stateless callers.
- High-volume, user-facing traffic, ramped in steps with your error and latency alerts watching.
- Tool-calling agents last, once the trace replay is clean.
Keep the old deployment alive but out of rotation until Sunday, so you have a rollback for any regression found Friday. After the date it will only return 410 Gone, so the rollback has a deadline, and the fallback target has to be a different deployment.
Do not use Nov 14–15 to finish work. Microsoft says migrating deliberately turns the retirement date into a formality;3 the weekend is for verifying that.
Sun Nov 15: verify, don't deploy
- Request logs show zero calls to the retired deployment names and ids.
- No 410 responses in application logs.
- Error rate and latency on the new deployments are flat against Friday.
- Anything still sending an old id has an owner and a ticket, not a hotfix at midnight.
A different failure on the date is the unexpected caller: a cron job, a notebook, an eval harness. Your Friday log search is the only way to find them before they find you.
Mon Nov 16 to Wed Nov 18: cut over the Nov 19 models
By Monday the o-series and Maverick replays should have results. Canary Monday, ramp Tuesday, finish Wednesday. Two points differ from the first track.
The silent swap is real for the o-series. Standard-type deployments on OnceCurrentVersionExpired (the behavior when the option is unset, per Microsoft's guidance as read) can be upgraded without an error.3 So a deployment you did not migrate will not look broken on Nov 19; it will look different. Either migrate it on purpose, or set NoAutoUpgrade and own the failure path. Auto-upgrades also roll out region by region, so two regions behind one route can answer with different models during the window.2
Provisioned o-series deployments get no safety net. Microsoft says replacement models become available in provisioned regions about 30 days before retirement, and that provisioned customers must migrate manually, in place (traffic moves over a 20–30 minute window) or side by side.2 3 Do side by side if you can keep both for a day.
Thu Nov 19 and after
- Zero requests on the old deployment names. Any 410 is a missed caller.
- No unexplained shift in output length or cost per request after the date, especially if a deployment was auto-upgraded.
- Delete the retired deployments and update runbooks, prompts and dashboards. Microsoft's Retire phase includes freeing capacity, archiving evals and updating downstream documentation.3
- Look at the next dates now. The same schedule lists gpt-chat-latest previews retiring 2026-11-17, claude-opus-4-5 on 2026-11-24 and claude-sonnet-4-5 on 2026-11-30 (listed replacement claude-sonnet-5-5).1 Cohere embed v3 models are deprecated on Nov 15, with retirement on 2027-02-15.1
Where a gateway helps
INFRO is a control plane between your application and the provider accounts you already hold; your providers bill you directly and INFRO takes 0% of model spend.4 For a cutover on this timetable, four shipped capabilities map to the checklist. Named routes with weighted targets let you ramp a canary without a client deploy; failure rules decide per failure kind whether to retry, move on or stop, so a retired deployment can fall to a named target (confirm in a drill how a 410 is classified before relying on it). Per-request usage by model shows which callers still send an old id, and a project model allowlist turns a straggler into a policy error instead of a late-night outage. Model retirements records vendor shutdown dates and refuses a retired model before any provider call, naming the replacement, but those are vendor dates, so keep the Azure dates above in your own registry. The INFRO docs (opens in a new tab) have the contract for each.
