MAI-Image-2.6 web_grounding flag: what Sume has instead

MAI-Image-2.6 and MAI-Image-2.6-Flash accept a boolean web_grounding parameter; when it is on, Microsoft says the model can retrieve current information from Bing Search and use it as context for the image (Microsoft Learn, read 2026-10-01). Sume's image API has no grounding or search switch on any model. If a picture depends on facts that change, such as a score, a price or a new product name, you put those facts in the prompt yourself.

Sume rejects fields a model does not list with 400 unsupported_parameter, so sending web_grounding fails loudly instead of being ignored. See the Image API reference for how capability descriptors work.

What each side offers

The table compares what is documented, not output quality.

How do I get current facts into a Sume image prompt?

Fetch the fact first, then state it as text the model should render or depict. Keep the instruction literal and short:

• Look up the fact in your own code or agent (a score, a date, a product name) and keep the exact string.
• Put it in quotes in the prompt: the sign should read "Final: 3-2", not "the latest score".
• Attach a reference image if the visual matters, such as the real product photo, through input_references.
• Generate, then check the text by eye before publishing; no image model is guaranteed to spell a string right.

When grounding actually helps

Grounding matters for real-world entities: a landmark, a current sports kit, a brand's latest packaging. Microsoft says it can improve accuracy for requests about real-world entities, places and events. If you need that on Sume today, a reference photo is the dependable equivalent, because the model sees the real thing instead of searching for it.

Google makes a similar claim for Nano Banana 2, which it says can use web image search (Build with Nano Banana 2, read 2026-10-01). Sume lists no search parameter on its Nano Banana rows, so the same workaround applies. See Nano Banana grounding on Sume.

Limits

A pasted fact fixes the words, not the layout, and long strings still misrender on some models. Microsoft's grounded results depend on what Bing returns at request time, so they can differ between runs; a pasted string will not. Both are previews of what the model can do, not guarantees. MAI is not in Sume's image list as of this post, so verify with GET /v1/images/models.