Is Gemini 3.8 Flash on Sume? An agent LLM, not a media model

Yes, Gemini 3.8 Flash is on Sume, but as an agent LLM: the model that reads a request and decides which tools to call. It is not a video or image model, so it never renders a frame. Sume's own repo describes the explicit auto selection on a Format run as Gemini 3.8 Flash first, then one GPT-6.1 Sol hop when the first model hits a classified transient failure, where the Auto catalog is enabled (dest first).

The question comes up because Google's launch list for September and October 2026 mixes three kinds of model: Gemini 3.8 Flash (a text and agent model), Gemini Omni Flash (video), and audio models. Only one of them is an LLM. This page separates them, using Google's changelog (read 2026-10-03) and Sume's repo and docs.

What Google says about Gemini 3.8 Flash

Google's Gemini API changelog lists Gemini 3.8 Flash on September 2, 2026 as its most intelligent Flash model, aimed at long-horizon software engineering, autonomous agents and complex enterprise workflows. The same changelog lists Gemini 3.8 Flash TTS and Flash-Lite TTS as generally available on September 22, and Gemini 3.8 Live on September 15.

None of those entries describes video or image output for 3.8 Flash. The video model in the same family is a different product line, Gemini Omni Flash, which the changelog lists as generally available on August 27, 2026.

What Sume's repo lists

On the Sume side there are two separate places where a Google model shows up. The agent model list in the provider proxy includes google/gemini-3.8-flash among its native OpenRouter ids, and the web app displays it as "Gemini 3.8 Flash". That is the LLM role.

Video is a different catalog. Sume's Video Router lists gemini-omni-flash-1.1 (Gemini Omni Flash 1.1), with 3 to 10 second clips at 360p, 720p, 1080p and 4K in 16:9 or 9:16, and native synced audio that is always on. There is no Gemini 3.8 Flash row in the video or image catalogs.

• Gemini 3.8 Flash: agent LLM, appears in the agent model list and in the auto orchestrator selection.
• Gemini Omni Flash 1.1: video model, id gemini-omni-flash-1.1 on POST /v1/video-router/generate and POST /v1/videos.
• Image models: a separate list (ChatGPT Image 2.5, Nano Banana 2 and Pro, Seedream, Flux 2, Ideogram and others), read it from GET /v1/images/models.

How the model field works on a Format run

The model field on a Format run selects the orchestrator only. Sume's Call a Format page says image, video and audio models are chosen by the Format's tools, that an id outside the Agents catalog is a 400 invalid_request, and that the receipt echoes the id that ran.

The OpenAPI description in the repo adds that an explicit auto opts into Gemini 3.8 Flash and then one Sol hop on a classified transient failure, and that concrete pins do not authorize model fallback. It also says served_model, fallback_used and model_attempts report what actually ran. Because that availability is described as enabled where the Auto catalog is on, with dest first, check the receipt on your own workspace instead of assuming it.

What Sume does and does not do here

If your goal is Google video, use the Omni row. If your goal is a Google LLM driving a workflow, use the orchestrator model field. For the function-calling angle see Gemini 3.8 Flash function calling.

• Does: let you pin an orchestrator id or send auto, and report the model that served the run on the receipt.
• Does: keep media model choice separate, so changing the LLM does not change which video model renders.
• Does not: render video or images with Gemini 3.8 Flash.
• Does not: promise that a given workspace has Auto enabled; the repo says dest first.