Gemini Omni Flash went GA Aug 27: five checks after a preview ends

Gemini Omni Flash became generally available on August 27, 2026, 58 days after its public preview began on June 30. The GA release added video extension, frame interpolation and a 360p to 4K resolution range, so any limit you measured during the preview is out of date.

Why a preview number is not a GA number

A preview is a moving target. The changelog shows the preview offering one duration band and one resolution, while the GA entry lists a wider resolution range and two new capabilities. Prompts, cost estimates and QA sheets built in June were measured against a narrower product.

The Sume catalog documents the GA model as gemini-omni-flash-1.1: 3 to 10 seconds at 360p, 720p, 1080p or 4K, in 16:9 or 9:16, with native synced audio. It accepts image and video references but not audio references, and its edit mode is reached through the video_url field on Video Router. Sume's docs do not describe extension for this model, so do not plan around it there.

The five checks

Run these once, on the GA id, and save the results next to your old preview notes.

• Duration: request the shortest and longest clip the catalog lists and confirm both come back at the length you asked for.
• Resolution: render one clip at each resolution you plan to ship and check the delivered pixel size, not just the request field.
• References: submit the maximum number of image and video references the catalog lists and confirm none are silently dropped.
• Audio: confirm the track is present on every output, since audio is always on for this model on Sume.
• Cost: compare the reserved amount at submit with the final usage.cost for each resolution, and rebuild your per-clip estimate from those numbers.

Read the limits from the catalog

The model list is the source of truth for supported durations, resolutions and reference types. Fetch it instead of copying numbers from a blog post, including this one.

curl -s https://api.sume.com/v1/videos/models \
  -H "Authorization: Bearer $SUME_API_KEY" \
  | jq '.data[] | select(.id == "gemini-omni-flash-1.1")
        | {supported_durations, supported_resolutions,
           supported_aspect_ratios, supported_input_references}'

What to do with old preview results

Do not delete preview outputs, but label them. A clip rendered on June 30 at 720p was made by the preview model, and a clip rendered after August 27 was made by the GA model. If you show both to a client as one batch, the difference in resolution options and capabilities can look like inconsistency in your own pipeline.

Rebuild any budget that used a preview price. Google's pricing page lists a GA price for the Omni Flash model, and Sume reserves at provider list times 1.25 on every model, so a sheet built from a preview-era number can be off in both places. Re-derive it from the reserved amount of a real GA job rather than from a table.

Finally, write the check results down with the date. The next model change then has a baseline to compare against, and a reviewer can see what was tested instead of taking the result on trust.

Keep the model id in config

Store the id in one config value and log the model field from each poll response next to the job id. When the next preview or GA date arrives you can find every output made on the old id and re-run only those. Job status and result reads are covered in Jobs and results.