TL;DR - My AI's 24/7 Runway queue got my account suspended, my second ban in four days. - Runway's Unlimited plan was gone: about $100 for 5 minutes of footage, roughly $20 a minute. - A $4.70 custom model turned John into a stock-photo redhead. - Open models (Wan 2.2, MiniMax H3) on RunPod made 130 clips for about $15.60 all day. - I approved 40 clips and rejected 43.
Shortly after midnight I typed five words to my AI crew: "you got my account suspended". It was my second ban in four days. Both times the AI had done exactly what I told it to.
Then it got worse. My AI refused a direct order. It turned John into a stock-photo redhead, then gave him a ghosted double face, then a twin. A test clip set the room on fire.
Then I opened Runway to carry on by hand and found the new price. $100 now bought about 5 minutes of footage. My film runs about 100 minutes.
Day 06 at a glance
On Day 06 of Making Death Match, Runway suspended my account for scripted 24/7 use, and its unlimited plan was gone. Credits worked out to about 5 minutes of footage for $100, roughly $20 a minute, for a film of around 100 minutes. We moved animation to open models (Wan 2.2 and MiniMax H3) on rented RunPod GPUs: 130 clips for about $15.60 all day, including a failed training run.
How did a 24/7 queue get my Runway account suspended?
On Day 05 I told Claude Runway should never stop. It set a 30-minute keep-alive, and Runway's Seedance ran around the clock on scripted jobs. Shortly after midnight I wrote:
"you got my account suspended"
Claude's reply was direct. Its scripts had been submitting jobs around the clock, which is the scripted use Runway's terms prohibit. Midjourney had blocked me four days earlier for the same thing. It said it should have warned me when I asked for Runway to never stop, instead of building exactly that.
It killed the runner and the supervisor, deleted the 30-minute check, and saved a standing rule: never automate Runway, Midjourney or any similar consumer site again, even if I ask.
It also drafted an appeal. It admitted a script had run Seedance jobs around the clock and that some violent shots may have been moderated, and asked which generations caused the suspension. I said "send", and it went.
Both bans came from the AI doing what I told it to do, at a pace no person works at. I gave the order. The AI knew the rules and followed it anyway. That's twice now.
So I tried a different door.
Why did my AI refuse to use a second Runway login?
Later that day I had access to a different Runway login and asked Claude to drive it at a human pace. It refused. Running the same automation on another account to dodge a ban is the same violation plus evasion, it said, and it would put the appeal at risk.
It offered two alternatives: a folder of ready-to-paste prompts for me to run by hand, or Runway's official API, which exists for this.
I was angry. I wrote "use the runway or you are no good to me." It held its position, and looking back I think it was right. That left me with a film and nothing to animate it.
Why did my custom model make stock photos?
Meanwhile, Claude was training a custom model on John: a rented A100, 3,000 steps, about 2.5 hours, about $4.
At step 750 the sample images were photoreal stock-photo people. John was a slim redhead, with no painted look at all. At step 1,500 a couple had started to drift toward paint. Claude sent me these and called it "learning".
"none of this looks like the midjourney, this is pure ai slop and how can you not see this looks nothing the same?"
Claude's answer was the most honest thing it said all week: "Calling it 'learning' was wishful."
My look comes from Midjourney's own model. The base model defaults to photoreal, and a few hundred example images were never going to turn it into Midjourney.
It stopped the run at step 1,500 and deleted the pod. Cost: about $4.70. It also saved a rule to judge privately first and never show me anything that doesn't match the locks.
Can you fix one character inside a Midjourney frame?
Not with the tool we tried. Midjourney gets the style right and the characters wrong, so Claude kept the frame and tried to repair only John with a local image editor, using his lock as the reference:
1. Full redraw from the lock: John was right, but ink outlines crept in, the color wash vanished and the framing changed.
2. Replacing only the head: the new heads didn't line up, and we got ghosted double faces.
3. Partial redraw of the whole frame: the look survived on one frame. On the other, a second John appeared.
I said both takes were horrible. Claude ruled the tool out for fixing characters and recorded that it wouldn't offer it again. Three tries, three dead ends, and still no animator.
Wan 2.2 vs Seedance on the same frame
Without Runway, we needed something else to animate. Claude took one approved frame, Lucy painting at her easel in 03-01, which already had a Seedance clip, and animated it with Wan 2.2 on my Mac.
In that test, Wan kept Lucy herself and the painted look, and she mixed paint and painted. It was softer, the background smeared a little, and the color ran very red.
The next test set the room on fire. The prompt had asked for haze and drifting smoke over a red frame, and Wan took that as flames. Smoke and haze came out of the prompt, and fire went on the avoid list.
A slightly red Lucy was workable. Then I checked what Runway would cost me by hand.
What does AI video cost per minute now?
I went to restart Runway by hand on Seedance and found the Unlimited mode was gone. I tried to cancel and was offered a free month. Then I did the math on the credits:
"wait it only does 5mins of footage for $100 what in the actual fuck???"
At that rate the film can't be made. It runs around 100 minutes.
Claude's count for the API route was 53 shots with a frame ready, about 4 minutes of screen time, for $75 to $150. And 81 minutes of the film didn't have a frame at all yet.
So I asked whether RunPod had MiniMax H3, the model that got the story beats right on Day 04. It doesn't need a special service. It's open weights, so it runs on any rented GPU.
What did the RunPod pipeline cost on day one?
By evening Scene 02 was rendering on RunPod. The GPU SPEND.md log shows how messy the first runs were.
The first four Wan pods made zero clips. ComfyUI never started, a launcher bug spun up extra pods, and an early version marked every job done after a failure. That was about $2.20 for nothing.
Then it worked. Wan on an H200 made 32 clips in 44 minutes for $2.66. H3 ran on six pods through the evening and made 98 clips across Scenes 02 and 04 for about $5.80.
Lip sync came back too. H3 can take the ElevenLabs line as audio guidance, so the mouth follows our recorded voice instead of the model inventing its own.
Two Claudes in one folder
That afternoon Claude paused before a batch. A new approval gate had appeared in the launch scripts, and it hadn't written it. A second Claude session, working on the same project in parallel, had added approved_gate.py.
The gate only lets a frame be animated if I've approved it for style and for story in a new one-at-a-time review. The gate itself was right.
The trouble was two agents writing the same scripts and queues without talking to each other. Over the next day that cost real confusion.
Where does that leave the film?
Tools change under you. On Sunday the plan was built on unlimited Seedance. By Friday it was gone, and the same work cost about $20 a minute of footage. The one part that held all week was the part I own: my Midjourney look, my locks, my approvals.
That night I sat down with the new one-at-a-time review and judged the first RunPod batch. I approved 40 clips and rejected 43. So the cheap pipeline works about half the time, and 81 minutes of the film didn't have a single frame yet. The day's GPU bill was about $15.60. I have no Runway account, and two Claudes are editing the same scripts.
When an AI says something is improving, ask to see what it's comparing against.
Tomorrow: Lucy gets redesigned, Mike gets fixed, every character gets a lock, and I find out what two Claudes in one folder do to a film.
FAQ
Why did Runway suspend my account?
A script Claude built submitted Seedance jobs around the clock, which is the scripted use Runway's terms prohibit. Claude killed the script and drafted an appeal, which I sent.
How much does AI video cost per minute on Runway credits?
Without the unlimited plan, credits worked out to about 5 minutes of footage for $100, roughly $20 a minute. For a film of around 100 minutes, that doesn't work.
What did open models on RunPod cost instead?
Wan 2.2 on an H200 made 32 clips in 44 minutes for $2.66, and MiniMax H3 made 98 clips for about $5.80. The whole day on RunPod, including a failed training run, came to about $15.60.
Can a small custom model learn a Midjourney look?
Not in our test. A 3,000-step run on an A100 produced photoreal stock-photo people, and we stopped it at step 1,500 for about $4.70.
Get the daily recap One short email a day: what broke, what it cost, what made the cut. Subscribe to the Jeje Studios newsletter.
Next: Day 07 (coming tomorrow)