The four retirements as one sequence
Microsoft's Foundry schedule lists four DeepSeek preview versions with retirement dates, and the replacement cell is empty for each1. Three dates matter, because two versions share December 15.
| Retires | Foundry preview version | Days from Nov 21 | Listed replacement |
|---|---|---|---|
| Thu Dec 3, 2026 | DeepSeek-V4-Flash-0731 | 12 | none |
| Tue Dec 15, 2026 | DeepSeek-V4-Pro-0813 | 24 | none |
| Tue Dec 15, 2026 | DeepSeek-V4.1-Flash (2026-09-10) | 24 | none |
| Sat Mar 6, 2027 | DeepSeek-V4.1-Flash-1017 | 105 | none |
Two rules from the lifecycle policy shape the plan. A retired model returns 410 Gone on every inference request. A preview deployment is either force-upgraded to a newer preview or to GA, or terminated with no replacement, and you cannot choose to stay on it2. Microsoft also says it does not recommend preview models in production3. Treat the dates as the day the call fails, and finish the day before. The pages say nothing about the time of day or time zone, and Mar 6 is a Saturday, so plan for Friday Mar 5.
Before the first move: inventory
Do this once, this week. Everything later depends on knowing which code calls which version.
- List every Foundry deployment whose model is one of the four versions above, with its owner and the services that call it.
- Record which calls use thinking mode, tools, JSON output, images or long outputs. DeepSeek's API documents thinking as the default mode and a 1M context with up to 384K output4; confirm what your Foundry deployment does with the same request.
- Pull a week of logged requests per workload into a test set, with the expected outputs or a rubric.
- Note whether finance relies on Azure billing, Azure support or an Azure spending commitment for these calls. Models sold by Azure are billed through your Azure subscription and covered by Microsoft support and Azure SLAs3; a move to DeepSeek's API leaves all three.
- Find hard-coded model or deployment names in code, prompts, notebooks and CI. These are the edits an alias removes.
Which replacement for which deployment
There are three destinations. Only the first two are confirmed.
| Destination | What is confirmed | What to check |
|---|---|---|
| GA on Foundry | DeepSeek-V4-Flash and DeepSeek-V4-Pro, both version 2026-04-23, GA, retiring 2028-02-221 | Version differs from your preview; vision and parameters |
| DeepSeek's own API | deepseek-flash serves V4.1-Flash-1017; deepseek-v4-pro serves V4-Pro-08134 | Billing, data handling, concurrency limits |
| SageMaker | Not confirmed (see below) | Whether a listing exists at all |
Dec 3: DeepSeek-V4-Flash-0731
This is the nearest deadline and the one with the least time to evaluate. DeepSeek describes 0731 as the same architecture and size as the earlier Flash preview, re-post-trained5. DeepSeek's own API no longer serves a V4 Flash: it was retired, and the old names are routed to V4.1-Flash-10175. So the like-for-like choice on Foundry is GA V4-Flash, which is a 2026-04-23 build and therefore older than 0731. Run your test set against both before you accept it. If the workload can take V4.1 behavior, deepseek-flash is the other candidate.
One oddity: in the schedule text I read, the replacement cell beside GA V4-Flash names the 0731 preview, which retires first. Do not follow that pointer, and recheck the live page.
Dec 15: DeepSeek-V4-Pro-0813
The two vendors' labels disagree here. DeepSeek's change log says the GA release of V4-Pro rolled out on August 135, and its pricing page lists deepseek-v4-pro as DeepSeek-V4-Pro-08134. Foundry lists 0813 as a preview and the 2026-04-23 version as GA1. Do not assume the Foundry GA row is the build you have been running. The DeepSeek API is the closest match by name; GA V4-Pro keeps you inside Azure. The choice is your eval result against your need for Azure billing and support. Mind the ceiling: DeepSeek lists a concurrency limit of 500 for Pro4.
Dec 15: DeepSeek-V4.1-Flash (2026-09-10)
Foundry's next preview, V4.1-Flash-1017, retires on Mar 6, so moving there only schedules a second cutover. GA V4-Flash is a different generation. DeepSeek presents V4.1 Flash as having native multimodal visual understanding5, and the pricing page marks vision as supported for deepseek-flash4. If your calls carry images, test GA V4-Flash on them before choosing it.
The cleaner option is deepseek-flash on DeepSeek's API. It already serves V4.1-Flash-10176, so the same move also covers the Mar 6 deadline.
Mar 6, 2027: DeepSeek-V4.1-Flash-1017
This one has the longest runway and the least certain end. Foundry says previews launch with a "not sooner than" date, typically 90 days out, and are sometimes extended until a replacement exists2. Treat the date as firm and check the schedule monthly for a GA V4.1 Flash or a later preview. If nothing appears by early February, deepseek-flash is the fallback.
Prices to compare, and what to leave out
DeepSeek's list rates, per 1M tokens in USD, as on its pricing page today4:
| Flash peak | Flash off-peak | Pro peak | Pro off-peak | |
|---|---|---|---|---|
| Input, cache hit | 0.006 | 0.003 | 0.044 | 0.022 |
| Input, cache miss | 0.30 | 0.15 | 1.32 | 0.66 |
| Output | 1.20 | 0.60 | 3.96 | 1.98 |
I did not read Azure's rates for the GA V4 deployments, so compare those from your own Azure price sheet and contract. DeepSeek deducts usage from a balance on your DeepSeek account4, which lands in different cost data than an Azure line item. Thinking is the default mode4, so measure output tokens on your own prompts rather than reusing rate-card arithmetic.
SageMaker: confirm before you plan on it
AWS's JumpStart overview names providers such as AI21 Labs, Cohere, Databricks, Hugging Face, Meta, Mistral AI, Stability AI and Alexa, and does not name DeepSeek7. I could not confirm a V4 Flash, V4 Pro or V4.1 Flash listing, so I would not build a December plan on one. If your cloud team has a verified model card, the trade-offs (instance-hours, capacity, data location) are in DeepSeek V4 on SageMaker JumpStart: self-host or call the API?. Also remember that a JumpStart build is a pinned set of weights, not whatever deepseek-flash serves next month.
Keep one alias through all four moves
DeepSeek's own names move under you. deepseek-flash was served by V4.1-Flash and now by V4.1-Flash-1017 with the calling method unchanged5. deepseek-v4-flash still works only because it is routed to another model6. The same applies to your own deployment names on Foundry. Put one name in application code and move the targets behind it. The wider argument is in model aliases vs pinned snapshots.
from openai import OpenAI
client = OpenAI(base_url="https://api.infro.io/v1", api_key=INFRO_API_KEY)
# "deepseek-flash-prod" is a route name, not a vendor model id.
resp = client.chat.completions.create(
model="deepseek-flash-prod",
messages=[{"role": "user", "content": "Summarize this ticket."}],
)If you run a gateway, this is what a route is for. An INFRO route is a name your application sends as model, resolved to an ordered list of your own provider connections and the model each should serve8. The weighted strategy picks the first target in proportion to weight and falls through the rest, and a weight of 0 is a standby that takes no first attempt, which suits a gradual migration8. The trace names the route and the connection that served the request8. INFRO takes 0% of model spend and your providers invoice you directly9. Confirm in the connection list that the provider kinds you need are served before you rely on this path.
Remove the old target before its retirement date. A 410 Gone is a client-side error, not a timeout or 5xx, and I did not confirm that failover covers it, so a leftover standby is not a safe net. INFRO also records vendor shutdown dates and refuses a retired model before any provider call, naming the replacement9; check the console to see whether these ids are in that record. To judge each target before shifting weight, build a dataset from logged requests and run it through an experiment, which shows an estimated cost and stops at a spend cap you set.
Cutover checklist
- Alias created, and every caller switched to it before any model changes.
- Eval run per workload on each candidate, comparing quality, latency and cost together, as Microsoft advises2.
- Dec 3 workload moved and the old target removed by Dec 2.
- Dec 15 workloads (Pro and V4.1 Flash) moved and removed by Dec 14.
- If V4.1 Flash moved to
deepseek-flash, Mar 6 workload marked done; otherwise a dated reminder on Feb 1 to re-read the schedule. - Billing owner told which spend now arrives in Azure and which on DeepSeek's balance.
- Concurrency checked against DeepSeek's limits: 2,500 for Flash and 500 for Pro4.
- Retired ids deleted from code, prompts and CI so nothing quietly calls them.
