Spent 2k Tokens While Trying to Make a 60-Second Video with Runway
I wanted to learn how AI video tools really work: how Runway and its models “think”, what they do well, where they fail, and how to write prompts that don’t waste money. I had never used Runway before.
To learn by doing, I picked a simple topic: two animated bear cubs run through the park holding hands. One trips on a stone and falls, and pulls the other one down with him. The message: if you run hand in hand and one of you falls, the other falls too.
It took a false start and two full tries. The false start cost about 300 Runway credits, the first try about 700 and the second about 1,000.
Tools
- ChatGPT
- Runway: Gen-4 Image Turbo, Gen-4 Image, Grok Imagine (image and video), Gen-4.5
- Claude Code
Before the first try
1. A false start (about 300 credits). I typed the story into Runway and expected the whole video in one go. The results were useless.
Note: A text prompt alone gives random results. Video models need pictures to start from.
2. Learning the right way. I watched some Runway tutorials and understood how it really works: first you create reference images, then you turn those images into video. So I made my bear cubs in ChatGPT.
Note: ChatGPT makes good-quality pictures cheaply, which makes it a good place to design your first characters.
First try (about 700 credits)
3. Pictures in Runway (spent 100 credits). I decided to continue in Runway, because it makes pictures too. I drafted every picture with Gen-4 Image Turbo (2 credits) and polished only the good ones with Gen-4 Image (8 credits).
Warning: Check how many pictures one click makes. If the setting is 4, every Generate makes 4 pictures and costs 4 times as much. Set it to 1.
4. The fifth scene picture. Both cubs and teacher in one picture failed 3 times in Turbo and 2 times in Gen-4 Image, because the model kept mixing them. I tried Grok Imagine, and it worked for just 2 credits. After that I used Grok for every scene picture.
Suggestion: If a model fails twice, don’t keep paying for the same failure. Try a different model. For pictures, Grok Imagine was the best and cheapest for me.
5. Ten clips (600 credits). I turned the ten pictures into ten clips with Gen-4.5, using prompts from Claude Code. Each try costs 60 credits. Here is the result:
Then I realized what was wrong: the reference images have to be really good, or the video looks weird.
Get Çağlar Can SARIKAYA’s stories in your inbox
Join Medium for free to get updates from this writer.
Note: A video can’t fix a bad picture. Before you animate, check every picture for the right pose, fingers and no stray letters. Pictures cost 2–8 credits, a clip costs 60–200.
Second try (about 1,000 credits)
1. New pictures. I made all the pictures again. Empty places like the playground or the path were easy. Pictures with the teacher and the kids together were hard and took many tries with different prompts.
Suggestion: Make the empty places first and reuse them. Expect most of your picture tries to go on scenes where characters interact.
2. The running clip. I kept only the first clip from the first try and remade the rest. The running shot was easy. I gave Runway the picture of the bears on the path and asked them to run.
In the first result they ran off onto the grass after two seconds. One more prompt, and they ran properly along the path.
Note: Describe what the characters keep doing, not where they go. A destination makes the model end the action there.
3. The fall (about 200–300 credits spent). This was extremely hard. I wanted the boy’s foot to catch on a stone while they run, so that both fall. I changed the prompt 4–5 times and made this clip 3 times in Gen-4.5. Then I tried splitting it into two parts, which didn’t work either. In the end Grok Imagine video did it, but it changed the characters’ style a little.
Warning: Accidents like trips and falls are among the hardest things for video models. Budget extra tries for them. Grok video may also add sound or voices, so mute it or listen carefully.
4. The teacher runs over (about 120 credits). My first prompt said the teacher saw the kids and came towards them, and the video came out silly. Then I prompted only this: the teacher looks surprised and worried and runs along the path. It worked.
Note: This makes sense, because Runway has no memory between clips. It doesn’t know the kids fell. Describe only what happens in this clip, in one simple action.
5. The teacher helps the kids (about 120–180 credits in Gen-4.5, plus Grok). The teacher reaches the kids, helps them stand up and separates their hands before they run again. She also shows with gestures that holding hands caused the fall. I tried 2–3 times in Gen-4.5 and finally got it with Grok.
Warning: Several characters doing several things in one clip means many tries. If you can, split it into one action per clip.
6. Running out of credits (the last ~200). By now I had about 200 credits left. In the next clip the kids ran backwards, but I accepted it. Then I asked for the kids arriving at the playground, and I gave Runway only this picture:
The clip came out nicely, and they arrived at the playground. But they weren’t bears anymore. They were two skunks 🫥
Warning: Without your character references, the model invents its own characters. Always attach them, or start from the last frame of a clip that has them.
With my last credits I made the final part. It wasn’t good, but that was everything I had.
Here is the result of the second try:
Total cost
About 300 credits on the false start, 700 on the first try and 1,000 on the second: about 2,000 Runway credits, around $35.
Suggestion: Keep a reserve. I ran out before the last scenes, which is where I had to accept mistakes.
