The timeline, in UTC and UTC+8
Alibaba's model page states the preview period runs from 10:00 on November 11, 2026 to 24:00 on November 25, 2026 (UTC+8), that no changes will be made during it, and that the model will be retired after the preview ends.1 Midnight at the end of Nov 25 in UTC+8 is 16:00 UTC on Nov 25. Set your own cutover deadline before that, not on it.
Model Studio's decommissioning policy gives preview models a shorter runway than stable ones: a notice no later than 14 days before End of Service, and End of Service 14 days after the start of retirement. Stable models get three months.2 Eleven days to twenty-five is fourteen days, so this window fits the preview rule (that is my reading; the policy page does not mention this model by name).
| Date | What the pages say |
|---|---|
| Nov 11, 10:00 UTC+8 | Preview opens, Token Plan only 1 |
| Nov 13 (today) | 12 days to the end of the window |
| Nov 25, 24:00 UTC+8 (16:00 UTC) | Preview ends; model retired after 1 |
| From End of Service | Inference stops, API calls fail, related console features and documentation are retired 2 |
The policy also says default rate limits may be gradually reduced once retirement starts, and that a retirement can be paused, postponed or cancelled.2 Plan for the stated date. Do not plan for an extension.
What "Token Plan only" changes
qwen4-max-preview is not available pay-as-you-go.1 That affects both the migration and the audit of who was calling it.
| Token Plan | Pay-as-you-go | |
|---|---|---|
| Billing | Subscription, consumed in plan quota units (Alibaba's name for them is on its pages) 3 | Per Model Studio's pay-as-you-go billing 2 |
| Endpoint | Dedicated https://token-plan.ap-southeast-1.maas.aliyuncs.com 1 | Workspace endpoint, https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1 4 |
| Key | Plan-specific key, shown once; Team seats get one key each 3 5 | Model Studio API key |
| Permitted use | Interactive use in compatible coding and agent tools; not automation scripts or application backends 6 | Not restricted that way on the pages read |
| Region | Singapore only, Global deployment, so inference runs globally and involves cross-border transfer 6 | Per workspace region |
| When quota runs out | Service pauses until the next cycle or until more quota is added 5 | Not applicable |
Three consequences follow.
- A plan key in a service is a policy problem. If any backend job, script or gateway connection used a Token Plan key for qwen4-max-preview, Alibaba's stated terms treat that as outside the permitted scope, with suspension or key revocation as the consequence.6 Do not carry that key forward to the replacement.
- The preview ran at a promotional rate. Alibaba says usage of this model consumed as little as 20% of the standard rate, and 8% between 22:00 and 08:00 (UTC+8).5 A pay-as-you-go replacement will not reproduce that curve.
- Billing names differ. The alias is
qwen4.0-max-preview. Both ids work, but bills show the alias.5 Search logs and invoices for both.
Find every caller before you change anything
Alibaba says it emails accounts that called a retiring model in the last three months, and points to the Model Monitoring page for the Singapore and China (Beijing) regions to see which retiring models your account uses.2 Start there, then check your own side. The notice may have gone to a mailbox nobody reads.
- Open Model Monitoring in each region you use and list calls to
qwen4-max-previewandqwen4.0-max-preview. - Grep code, config, prompt-management tools and agent settings for both ids, including editor and coding-agent configs that point at the plan endpoint.
- Group your request log by model id and key for the last 30 days. Note which keys are plan keys.
- Flag any plan key used from a server, scheduler or CI job.
- Record per caller: modalities (text only, or images and video), tool use, structured output,
reasoning_effort, and themax_tokensthey send. - Name an owner per caller and a cutover date earlier than Nov 25.
Fallback order
Alibaba's pages I read do not name a replacement for qwen4-max-preview. The Model Studio overview lists Qwen3.7-Plus and Qwen3.8-Flash beside it, and its quickstart calls qwen3.8-max on the workspace endpoint.4 The Token Plan tables list qwen3.8-max, qwen3.7-max and qwen3.7-plus as reasoning models; qwen3.7-max is marked text-only, while qwen3.8-max and qwen3.7-plus also take vision input.5 That is enough to propose an order. Treat it as a hypothesis to test.
qwen3.8-max, the closest tier. Primary candidate for everything, including image and video inputs.qwen3.7-max, text-only traffic only. Skip it for any route that sends images.qwen3.7-plus, for vision traffic if step 1 is unavailable or over a rate limit.- A different vendor's flagship, on a connection you already hold, for outages and policy blocks. Test it as a fallback-equivalence case, not an assumption; the earlier post on testing fallback equivalence covers the method.
Confirm each id is enabled for your pay-as-you-go workspace and region before it goes into a route. A model that appears in a Token Plan table is not proof it is open to your workspace.
Parameters that behave differently
Starting with qwen4-max-preview, max_tokens on Chat Completions equals max_completion_tokens and caps thinking and the final answer together. Existing Qwen models keep the earlier behavior, where the limit varies by model and thinking_budget.7 Thinking is on by default, reasoning_effort takes low, medium or xhigh (default), and it should not be set together with thinking_budget.1 Chat Completions also enables preserve_thinking by default, so return prior thinking in reasoning_content, not inside content.1
So one limit cannot be copied across the chain. Keep per-model settings and replay the same prompts against each.
from openai import OpenAI
# Point at your gateway or at the pay-as-you-go workspace endpoint.
client = OpenAI(api_key="...", base_url="https://api.infro.io/v1")
CHAIN = [
# model id, extra params that must not be shared across models
("qwen3.8-max", {"max_tokens": 4096}),
("qwen3.7-max", {"max_tokens": 4096}),
("qwen3.7-plus", {"max_tokens": 4096}),
]
def run(prompt: str, model: str, params: dict):
r = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
**params,
)
c = r.choices[0]
return c.finish_reason, r.usage.completion_tokens, c.message.contentFlag finish_reason == "length" as a failure, and compare completion-token counts per prompt against what qwen4-max-preview produced. Answer length is where the thinking-budget change shows up first.
Cutover checklist
- Build a replay set from real prompts per caller: at least the tool-calling, structured-output and image cases.
- Run every model in the chain on that set; record schema-valid rate,
lengthstops, latency and tokens per request. - Set the per-model output limit and reasoning setting from the replay results, not from the old values.
- Create pay-as-you-go keys for backend callers and revoke the plan key from any server that held it.
- Re-estimate monthly spend from measured tokens at the list rates in your own pay-as-you-go account, not from plan consumption.
- Ship the new chain to a small share of traffic first, with the old model still serving the rest, and watch error mix and truncation.
- Move all traffic by about Nov 20, leaving five days of margin, then remove
qwen4-max-previewfrom configs. - After Nov 25 (UTC+8), search logs for any remaining calls to either id.
Where a gateway helps
A gateway does not make a retired model answer. What it can do is keep the model id out of application code. If a Model Studio pay-as-you-go endpoint is connected as an OpenAI-compatible connection,8 a named route can hold the chain above as priority targets over your own connection, with failure rules deciding retry, move on or stop per failure kind. INFRO's failover retries a route that times out or returns a 5xx before the response starts. A retirement error is not that case, so remove the retired model from the route before the date, not after. INFRO also records vendor deprecation dates and, from the shutdown date, refuses a retiring model before any provider call and names the replacement. Check in the console whether this model is recorded. You still need the replay results above to choose the order.
