Which Claude retirements matter on Foundry
This checklist is for teams that call Claude through Microsoft Foundry because the spend counts against an Azure commitment, because prompts must stay inside a defined boundary, or both. Generic Azure retirement posts cover the OpenAI rows; the Claude rows behave differently because Anthropic sells and operates the models and Foundry offers them in two hosting versions.1
The Claude rows below come from Microsoft's retirement schedule, which was last updated on 2026-09-23, so confirm against your own subscription before acting.2
| Foundry model id | Hosting | Retires | Replacement listed |
|---|---|---|---|
| claude-sonnet-4-5 | Anthropic | 2026-11-30 | claude-sonnet-5-5 |
| claude-opus-5-5-fast | Azure | 2027-01-15 | none |
| claude-opus-4-6 | Anthropic | 2027-02-02 | none |
| claude-sonnet-4-6 | Anthropic | 2027-02-10 | none |
| claude-opus-4-5 | Anthropic | 2027-02-22 | claude-opus-5-5 |
Two cautions. First, the brief for this post named a Nov 24 date; no Claude row on the schedule I read carries it, so it is not in the table. Second, the schedule's own Replacement column is Microsoft's recommended target and "doesn't limit your manual migration choices", and Microsoft declares it roughly 90–120 days before retirement, which is why three rows are still blank.4 For Sonnet 4.5, see the Sonnet 4.5 test method.
Check what you actually run
- List every Foundry Claude deployment per subscription, with model id, hosting version and deployment type. The schedule keys rows by model and version, and an "existing customer" is decided per subscription.4
- Mark each as Azure-hosted or Anthropic-hosted. Opus 4.6, Sonnet 4.6 and Opus 4.5 are listed only as hosted on Anthropic infrastructure; Opus 5.5 Fast only as hosted on Azure.3
- Record the retirement date, and subtract your validation and rollout time. Microsoft's own advice is to treat that difference as the real deadline.5
- Confirm someone receives the notices. GA retirement notices go out at least 60 days before retirement, by email to subscription owners with active deployments and through Azure Service Health.4 By that rule the notice for a Jan 15 retirement is due around Nov 16 (my arithmetic).
- Pull lifecycle fields from the Models API, which exposes
lifecycleStatusand per-SKUdeprecationDate.4
TOKEN=$(az account get-access-token --query accessToken -o tsv)
curl -s -H "Authorization: Bearer $TOKEN" \
"https://management.azure.com/subscriptions/$SUB/providers/Microsoft.CognitiveServices/locations/$LOC/models?api-version=2024-10-01"Filter the output for your Claude ids. The page describes the API as covering any model; verify that Claude rows appear for your region before you rely on it.
Decide what the boundary actually requires
Moving off a retiring Anthropic-hosted model is also a chance to change where prompts are processed, so settle the boundary question first.
According to Microsoft, Azure-hosted Claude processes prompts and outputs on Azure infrastructure end to end, with data at rest in the selected geography. Anthropic-hosted Claude processes data on Anthropic's infrastructure, possibly outside Azure and outside your selected region.1 Both versions list Anthropic as seller, operator and data processor, and both bill in Claude Consumption Units that decrement an Azure Consumption Commitment.1
That gives three situations:
- Commit only. Either hosting version satisfies it, so pick on features and quota.
- Boundary matters. Only Azure-hosted counts, and the deployment type matters: Azure-hosted offers Global Standard and Data Zone Standard (US); Anthropic-hosted offers Global Standard only.1
- Boundary matters and you are outside the US. Microsoft's comparison lists no regional data zone other than the US one for Claude, so check whether Global Standard fits your policy before you plan around Foundry alone.
Also check the default: if a model exists in both versions, a new deployment lands on the Azure-hosted version.1 That is probably what you want, but it is a change your security reviewer should see.
Replacements that stay on Foundry
Microsoft lists these Claude models as GA on Foundry, with their hosting versions:3
| Replacing | Candidate | Azure-hosted | Anthropic-hosted | Retires |
|---|---|---|---|---|
| Opus 4.5, Opus 4.6 | claude-opus-5-5 | Yes | Yes | not listed on the schedule page; check before committing |
| Opus 5.5 Fast | claude-opus-5-5 | Yes | Yes | not listed on the schedule page; check before committing |
| Sonnet 4.6 | claude-sonnet-5-5 | Yes | Yes | not listed on the schedule page; check before committing |
| Sonnet 4.6 | claude-sonnet-5 | Yes | Yes | 2027-06-30 |
| Opus 4.6 | claude-opus-4-8 | Yes | Yes | 2027-09-01 |
Opus 4.5 is the only retiring row whose replacement Microsoft names, and it names claude-opus-5-5.2 For Opus 4.6 and Sonnet 4.6, choosing the 5.5 generation is my recommendation, not Microsoft's. Prefer the newest GA model that is not itself close to retirement: Anthropic GA models on Foundry follow a 12-month lifecycle rather than the standard 18 months, so an older "safe" target ages out sooner.4
Opus 5.5 Fast is the odd row: it is Azure-hosted only, retires 2027-01-15, and has no listed replacement. The nearest same-hosting move is claude-opus-5-5, which is also Azure-hosted; you give up whatever the Fast variant provided, so re-measure latency.3
What breaks when you change versions
- Thinking parameters. Opus 5.5 always runs adaptive thinking; requests that set thinking to disabled, or to enabled with a manual
budget_tokens, are rejected with a 400.3 Opus 4.5 is listed with extended thinking, so its callers are the most likely to hit this. - API surface on Azure-hosted. Azure-hosted Claude supports Messages and token counting;
/filesand/skillsare listed only for Anthropic-hosted.1 Search your code for both. - Context and output limits. Opus 4.5 is listed at 200K context and 64K output; Opus 5.5 at 1M and 128K.3 Prompt-size routing rules and truncation guards need new numbers.
- Tokenizer and cost. A new generation can change token counts for the same text; see how a tokenizer change moves your bill.
- Behavior. Replay frozen production inputs on the target unchanged before tuning prompts, tools or schemas, as Microsoft's six-phase process advises.5 Capture is opt-in and never retroactive, so turn it on now.
Auto-upgrade is not a plan. Microsoft describes it as a safety net for pay-as-you-go deployment types, controlled by versionUpgradeOption, and notes that provisioned deployments get none.5 Whether it applies to a partner Claude row with no listed replacement is not stated on the pages I read, so assume the deployment stops and returns 410 Gone.4
A failover order that respects the boundary
Retirement day is one failure; a regional quota or capacity problem is another. A fallback chain should only ever step down a constraint on purpose. Order by what you must keep:
- Primary: the Azure-hosted replacement on your chosen deployment type.
- Same model, second Azure deployment: another region or the other deployment type, within what your policy allows. No change in model, hosting or commit.
- Previous generation, Azure-hosted: for example claude-opus-4-8 for an Opus 5.5 primary. Same boundary, with a retirement date of 2027-09-01.2
- Anthropic-hosted Claude on Foundry: same commit, same invoice, but data may be processed outside Azure.1 Include it only if your policy permits.
- Non-Claude on Foundry: models sold by Azure are hosted and operated by Azure, billed through your Azure subscription and covered by Azure service-level agreements.6 Foundry lists gpt-6.1-sol and gpt-6.1-terra there, among others.6 If your boundary is strict, move this above step 4. It needs its own quality test.
- Outside Foundry: Claude from the Anthropic API, Bedrock or Google Cloud. This leaves both the Azure commit and the Azure boundary, so it needs sign-off, not just a config change. See which commit each cloud burns.
Before you wire step 5 or 6, run a fallback equivalence test so a failover does not quietly change answer quality.
Where a gateway helps, and where it does not
A gateway cannot move the boundary for you, but it can keep a retired id from reaching production and keep the chain above enforceable. INFRO records vendors' retirement dates, warns organizations still sending a retiring model, and from the shutdown date refuses it before any provider call, naming the replacement. It retries a request that times out or 5xxs before the response starts on another target, tries fallbacks models in order (up to five), and lets a model allowlist block any model a project is not approved to call.7
One limit: fallbacks run over your own connections for that model's vendor, so the set of connections is the boundary. INFRO's provider page lists Anthropic, Google Vertex AI, AWS Bedrock and Azure OpenAI among the accounts you can connect and does not, on the page I read, list Foundry-hosted Claude.8 Check the providers list (opens in a new tab) before planning a Foundry leg through INFRO, and do not connect an account whose hosting your policy excludes.
Cutover checklist by date
- Now: inventory deployments, capture traffic, decide hosting version, confirm notice recipients.
- Before 2026-11-30: Sonnet 4.5 callers move to claude-sonnet-5-5.2
- By mid-December: replay and score the Opus 5.5 Fast move; deploy the target side by side so rollback stays possible.5
- Before 2027-01-15: Opus 5.5 Fast traffic fully on its target; the old deployment deleted.
- Before 2027-02-02: Opus 4.6 traffic moved, thinking parameters fixed,
/filesand/skillsuse resolved. - Before 2027-02-10: Sonnet 4.6 traffic moved and limits updated.
- Before 2027-02-22: Opus 4.5 traffic moved; Haiku 4.5 moves to claude-haiku-5-5 on the same date.2
- After each date: confirm zero calls to the retired id and that the failover chain still ends where your policy says.
