Genspark launches Media Studio, a video agent workbench spanning 70+ model versions from 20+ vendors

Genspark has officially released Genspark Media Studio, a video agent workbench that puts 70+ model versions from 20+ vendors on a single canvas. The details come from a hands-on test published today by QbitAI (量子位), a Chinese-language outlet, which also relays Genspark's own description of how the product works.

Model lineup

An LLM serves as the "director" and offers three selectable tiers: DeepSeek V4.1 Flash, Claude Opus 5.5 and Claude Fable 5.1.

The models doing the production work include:

• Video: Seedance 2.5, Kling, Veo
• Image: GPT Image, Nano Banana
• Audio: ElevenLabs and other audio models (MiniMax TTS was used in one test)

According to Genspark, the product covers video, image, audio and 3D generation.

Promotion

From 18:00 Beijing time on October 27 to 18:00 on November 4, Plus and Pro members pay half the usual credits in Media Studio. The discount covers planning, image, video, audio and 3D generation. Credit pricing is not yet disclosed in the article.

How it works, according to Genspark

• Director LLM: It interprets the request, writes the script and assigns tasks to other models.
• Skills: Production workflows are packaged as skills. Ads and narrative shorts follow different ones, and the agent picks by task type.
• Story Bible: Character appearance, costumes and other settings are stored as fixed reference assets. They can be reused in other projects.
• Automatic checking: The agent compares generated clips with the script and settings and re-shoots those that fail.
• Reshoot clips: Users mark a time range on the timeline and describe the change. Genspark says the original motion and camera position are kept. Up to 5 segments can be specified per request, each at least 4 seconds.
• Frame edit: Users can replace or remove an object in a video of 30 seconds or less.
• Code-to-video: Knowledge Explainer mainly renders video from AI-written code. Genspark says code rendering itself costs no credits, while narration and a few atmosphere shots are billed separately.
• 3D whitebox previs: The agent lays out character positions and camera angles with whiteboxes before any video model is called.
• Export: Professional users get a layer editor and multi-track timeline, with export to PSD, Premiere projects or CapCut drafts.

QbitAI's test results

Rap video. QbitAI used a template and uploaded photos of two real people. Uploading real-person photos was blocked by the video model, and Genspark prompted a copyright check. After the tester chose a semi-realistic style, the agent called Seedance 2.5. The page estimated 8-14 minutes, but QbitAI says it felt like two to three minutes. The first version had no backing track, so the tester had ElevenLabs generate an original rap instrumental and ran lip sync. Initial drafts render at 480p, with an upgrade to 1080p after approval.

Knowledge explainer. QbitAI generated a video about a film with no uploaded material. The first version had flat narration and slideshow-like visuals. After a revision request and an ElevenLabs soundtrack, the tester judged the narration and music acceptable, though the visuals still resembled a slideshow.

Short drama. The Story Bible feature built a project from a title alone. The tester approved 35 items, including character settings, costumes, props and styled portraits. QbitAI notes some approval options were hard to find, and any edit sent already-approved items back to pending. The agent split 5 scenes into 84 shots and 12 video units, generated clips with Seedance 2.5, repaired mismatches itself, then edited, scored and added bilingual Chinese-English subtitles. Flaws included audio-video desync in the second half, frame skips, and a character's glasses vanishing.

Game. A pixel-style HTML Mario-like game came out quickly with stomping, mushrooms, brick bumping and flagpole scoring. It cost only a few hundred credits. The three video cases used most of the starting credit allotment.

What this means

Media Studio's main claim is workflow consolidation: scripting, storyboarding, character settings, generation, repair, editing, music and subtitles run in one place. QbitAI's test shows the agent handling most of those steps, but it also records limits: audio-video desync, frame skips, continuity errors, and approval friction. The 480p draft stage, whitebox previs and code-to-video are credit-saving design choices, since problems can be caught before full-quality rendering. Per-credit costs outside the half-price window are not stated.