Hy Image 3.5 Preview dated id 20260922: should you pin the snapshot?

Pin the dated id if a client has approved how your images look, and float to the undated name only if you want the vendor's changes to reach you automatically. The OpenRouter listing for Tencent Hy Image 3.5 Preview shows the id tencent/hy-image-v3.5-preview-20260922 alongside the base slug tencent/hy-image-v3.5-preview.

This is a preview, which is a reason to lean toward pinning: a pinned id is your best chance of rerunning last week's brief and getting the same family of results.

What the id tells you

The listing, read on 2026-10-07, shows the dated id above and a released date of Oct 5, 2026; the page does not explain the 20260922 suffix. It does not publish a deprecation date, a changelog or a promise about how long the dated id stays available, so none is assumed here.

A dated id usually names one snapshot of a model, while a name without a date can move. Treat that as a convention to verify, not a guarantee: check the vendor page before you rely on it for a contract.

Where Sume fits

Sume does not list a Hy model in its image catalog, so you cannot call it through POST /v1/images today. For models Sume does list, the id you send is the catalog id, and GET /v1/images/models shows the current rows; the model field in the response echoes what you asked for.

Sume keeps this simple for the ChatGPT Image 2.5 family: openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst are separate ids with the same token rates, so you choose a lane explicitly. See the Image API docs for the details.

• Write the model id into your job record, not just the prompt.
• Re-run a three-prompt test whenever the vendor page changes.
• Keep the previous dated id in your notes until the new one passes.

Testing a preview before you commit

A preview model deserves a small acceptance test, not a leap of faith. Pick five prompts that match your real work, two of them with references, and run each three times. Save the outputs next to the dated id and the date. If the three runs of one prompt vary a lot, a pinned snapshot gives you less repeatability than you may have hoped, and that is worth knowing before a client sees it.

The listing describes multi-turn editing, so test that path too. Ask for a change, then a second change on the result, and check whether the first change survives. A model that drifts across turns is a poor fit for an approval loop even if single images look strong.

Record the id, the date you tested, the size and the number of references for every approved image, and store the model id in the same record as the job result. On Sume, the model field of the response echoes the id you requested.

• Keep the test prompts in version control with the id.
• Re-run the same five prompts when the vendor page changes.
• Do not generalize from one lucky output.