Script → storyboard → first cut
Give every scene the same starting point.
Creative Claw is a ChatGPT app that turns a written script into a multi-scene short film with the same character in every shot. Save a reusable Character, approve a shot-by-shot storyboard, render each scene with a reference-capable model such as Seedance 2.5 or MiniMax H3 Max, and assemble the shots into a first cut. References guide consistency; they do not guarantee an identical face or outfit in every frame. OpenAI discontinued Sora in 2026, so ChatGPT has no built-in video generator; you need an app like this one.
Start with 100 free credits. No card required.
By Itay Dressler, founder of Creative Claw · LinkedIn · Updated
Watch the 52-second tutorial
Bring
A script, character references, and a target film length
Create
An ordered first cut assembled from approved rendered shots
Finished example
“The Keeper” is an original short made for this guide. We saved Ada as a reusable Character from one reference sheet, generated a start frame for each scene, then rendered all four scenes in a single 15-second MiniMax H3 Max generation, using the character sheet as the identity reference and the four frames as scene references. The narration was generated separately with ElevenLabs and placed on each scene, with captions timed from its word timestamps.
Voiceover script
“Every night for forty years, Ada has climbed these stairs. The sea doesn't know her name. But tonight... someone out there needs the light. And every morning... it's worth it.”
| Step | Tool | Credits |
|---|---|---|
| Character reference sheet | Nano Banana 2 | 16 |
| Four scene frames from that sheet | Nano Banana 2 | 64 |
| All four scenes in one 15-second generation, 1080p | MiniMax H3 Max | 485 |
| 10.7-second narration | ElevenLabs v3 (Foley) | 4 |
| Place each line on its scene and caption it | ElevenLabs alignment | 0 |
| Total for the finished example | 569 ($5.69) | |
Charged costs from this single run on September 27, 2026. H3 Max renders up to 15 seconds with hard cuts in one generation; for longer films, render scenes separately and assemble them. Identity is reference-guided, so review every scene.
Measured example
Compare unedited Seedance 2.5, MiniMax H3 Max, and Wan 3.0 results using the same character sheet, ElevenLabs voice reference, and controlled three-scene prompt.
The workflow
Connect Creative Claw in ChatGPT. Supply a reference you have permission to use, or create a fictional character sheet. Settle the face, hair, outfit, and visual style, then save the approved reference as a reusable Character.
Create a Film project and a storyboard. Give each shot a purpose, action, camera direction, duration, and reference requirements. Approve the plan before rendering; account for the chosen model’s supported clip lengths.
Ask for a model that supports the reference inputs your shots need. Reuse the approved character reference across applicable shots, not just the character name. Review model parameters and a per-shot estimate before generation.
Generate the shot clips and inspect identity, wardrobe, product details, motion, and continuity. Resolve missing or failed clips and approve any retries. A saved Character is not a guarantee that all generated frames will match.
Once every required shot has a rendered clip, assemble them in project order. Add the project narration if appropriate, then review the result. Captioning, detailed transitions, and per-shot sound design may require separate finishing work.
Pricing
Creative Claw is pay as you go: $10 buys 1,000 credits, new accounts get 100 free credits, and there is no subscription. Typical costs for this workflow:
| What you generate | Credits | Cost |
|---|---|---|
| Character reference sheet, Nano Banana 2 | 16 | $0.16 |
| Storyboard frame per scene, Nano Banana 2 | 16 | $0.16 |
| One 15-second multi-scene generation, MiniMax H3 Max (1080p) | 485 | $4.85 |
| One 5-second clip, Gemini Omni Flash (720p) | 75 | $0.75 |
| One 5-second clip, Seedance 2.5 (720p) | 267 | $2.67 |
| A 10–15 second ElevenLabs v4 voiceover | 4 | $0.04 |
| The finished 15-second film above | 569 | $5.69 |
Prices charged or quoted by Creative Claw on October 18, 2026. See all model prices, and ask for an estimate in ChatGPT before you generate.
Alternatives
The tools below are the ones AI assistants most often recommend for this job. We checked each vendor’s own pages on September 27, 2026 (Higgsfield and Runway on October 18, 2026). “None found” means we could not find an official app in the ChatGPT directory or on the vendor’s site. Creative Claw publishes this page.
| Tool | Official ChatGPT app | Consistent characters | Video models | Cheapest paid plan |
|---|---|---|---|---|
| Creative Claw | Yes | Reusable Characters as references | Several: Seedance 2.5, MiniMax H3 Max, Gemini Omni and more | Pay as you go: $10 = 1,000 credits, no subscription |
| Higgsfield | Yes | Soul ID | Several: Seedance 2.5, Kling 3.0 and more | $19/mo Starter, selected models only |
| Runway | Yes | Gen-4 References | Several: Gen-4.5, Kling 3.0, Veo 3.1 | Pro $35/mo ($28 billed yearly) for MCP |
| Sora 2 (OpenAI) | Discontinued | Character assets (API only) | Sora 2 only | Shut down: app April 26, API September 24, 2026 |
| Vidu | None found | Reference to Video, up to 7 references | Vidu models only | $10/mo ($8 billed yearly) |
| LTX Studio | None found | Elements, from the Standard plan | Third-party models from the Standard plan | $15/mo ($12 billed yearly) |
Sources: Higgsfield pricing · Higgsfield in ChatGPT · Runway pricing · Runway Gen-4 References · Sora discontinuation · Vidu pricing · Vidu Reference to Video · LTX Studio pricing. Prices are in USD before tax and change often.
Before you render
Identity references and first-frame inputs have different purposes. Ask for a model and input mode that support the desired scene composition; forcing the same starting frame into every shot can limit the storytelling.
Film assembly concatenates rendered shots and can use project narration. It does not automatically provide full transitions, per-shot audio mixing, or captions. Keep those deliverables explicit.
Test a short scene before a long film. Budget for reference images, every shot, narration, finishing, and any approved retries. Longer projects amplify both cost and consistency challenges.
Check current pricing, read short answers in the help center, inspect the tool reference, or browse published media examples. Examples illustrate creative approaches; they are not proof that this exact starter prompt was run.
Practical answers
Not anymore. OpenAI shut down the Sora app on April 26, 2026 and the Sora 2 API on September 24, 2026, and ChatGPT does not generate video by itself. To make a multi-scene video with a recurring character in ChatGPT, use an app such as Creative Claw: save the character once, then render the scenes with Seedance 2.5 or MiniMax H3 Max.
Reusable character references help guide appearance across shots when the selected model supports them. Faces, outfits, and proportions can still drift. Approve a character sheet and inspect each rendered scene before assembly.
Creative Claw can. Paste the script into ChatGPT, save the main character as a reusable Character, approve a storyboard, then either render all scenes in one generation (MiniMax H3 Max handles up to 15 seconds with hard cuts) or render scenes separately and assemble them. Add narration and captions last. The Keeper example on this page was made in a single H3 Max generation.
The four-scene, 15-second example on this page cost 569 credits ($5.69): 80 for the character sheet and scene frames, 485 for one MiniMax H3 Max generation at 1080p, and 4 for narration. Longer films add about 16 credits per scene frame plus the video time you render.
Yes. Generate narration with an ElevenLabs voice and place each line on its scene, or describe dialogue and sound in the shot prompt for models with native audio. Review lip sync and timing before sharing.
No. The saved Character supplies a reusable description and reference. Generation must use appropriate reference inputs, and the model may still change details. It is reference-guided consistency, not a guaranteed identity lock.
Yes. Plan several shots within the model’s supported durations, render them separately, and assemble their completed clips. The total film length comes from the edit, not a single model call.
No. The project needs completed shot clips before assembly. Generate or resolve the missing clips first, then assemble the first cut.
Film assembly can use the project’s narration track, but it is not a complete post-production editor. Request captioning and other supported finishing work separately, and review sound and timing before sharing.
Start creating
Connect once, get 100 free credits with no card required, and use the starter prompt above for your first project.
