A legacy window against no window at all
One entry publishes a legacy window measured in months, with a clause that can close it after a fortnight of disuse. The other has no window at all: a copy already downloaded keeps working, so there is nothing for a date to be early against. As of 2026-09-22.
| On this point | Amazon Bedrock | Wan-AI |
|---|---|---|
| Is there a window | Yes, six months or 45 days | No; nothing is scheduled to stop |
| Can it end sooner | After 15 days of inactivity | Not applicable |
| What ends on a date | Access, in every region | Nothing you already hold |
| What ends without a date | New adoption, once legacy begins | New downloads, if the listing goes |
| What replaces the risk | A migration you perform | Hardware, storage and operations |
Inclusion rule. Both cells on a row come from the entry's own published pages: a lifecycle page for the platform, a repository listing and model card for the publisher. Order. Whether a window exists first, then what ends and when.
1Two ways to make timing a non-issue
One entry manages timing by publishing a window and a set of states. The other removes timing from the problem by handing over the artefact. Both are defensible and they demand different work from a production.
The first demands attention on a schedule. The second demands infrastructure continuously, and never demands attention at all.
2The inactivity clause has no counterpart
Nothing about holding weights can be shortened by not using them. That is worth stating because the platform's clause is the one sentence in this register that punishes the normal rhythm of serial production.
A team that pauses between episode blocks is exposed on one arrangement and untouched on the other. The reading for the clause sets out why fifteen days is an ordinary gap.
3Procurement risk replaces schedule risk
If a repository listing comes down, nobody loses a running copy and anybody who has not fetched it yet loses the chance. That is a purchasing problem rather than a lifecycle one, and it arrives without a date too.
The mitigation is keeping the weights rather than the address. Nothing publishes a state that will remind a team to do it, which is the trade recorded across every reading for that entry.
- The three states a model can be inActive, Legacy or End-of-Life, readable from the modelLifecycle field
- Models publishedWan2.2-T2V-A14B, Wan2.2-I2V-A14B, Wan2.2-TI2V-5B, Wan2.2-S2V-14B and the Animate variantsweights rather than an endpoint
4Sources
Both readings come from the pages the entries publish themselves: Amazon Bedrock and Wan-AI, read 2026-09-22. The column being compared is Ending early. Other pairs: Hailuo AI and Pika, SceneMixer and Pika.