The Last Subscriber: an AI founder-drama series, from 112 views to almost none
A five-episode AI-generated drama with a recurring character: credits per render, the defects we caught, and the view counts with their read dates. Run 2026-10-23; numbers read up to 2026-10-28.
Format: Serial drama, first-person narration, cliffhanger endings, five episodes, under a minute each, vertical 9:16. Everything shown in the videos is AI-generated. Numbers are from our own accounts and are small; each carries the date it was read, because platform numbers change as a video ages. This page makes no promise of views.
What we made
The Last Subscriber is a short AI-generated drama. A laid-off engineer builds an app alone; for a year exactly one anonymous user pays for it. Each episode ends on a cliffhanger. Characters and companies are fictional and original, adult, with no real people or brands and no income claims. Every post is labelled as AI-generated.
It was made the same way as the Dumpling series: a character sheet for the heroine so her look is pinned, a season plan with a price table, a free quote per episode, then the create call. It went through AutoShort's series API, not an MCP chat, so read the credits as what a character series costs, not as an MCP transcript.
The first attempts failed our own checks: a laptop in shot 1 carried a brand logo although the prompt said no brand mark, and the heroine's sheet had her holding a phone, which the pipeline copied into several shots. We rewrote the plan and the sheet and rendered again. Episode 1 was rendered four times before we posted it.
Credits and cost per render
| What | Number | As of | Source file : line |
|---|---|---|---|
| Episode 1, first test render (3 scenes, not posted) | 290 credits | run 2026-10-23 | README.md:19 |
| Character sheet for the heroine: two of three attempts failed and were refunded automatically | 21 credits refunded per failed attempt | run 2026-10-23 | README.md:17 |
| Episode 1, plan v1 (fixed after QA: logo, extra phone) | 328 credits | run 2026-10-23 | qa-ep1.md:10 |
| Episode 1, plan v2 (still had a logo on a laptop) | 328 credits | run 2026-10-23 | qa-ep1-v2.md:13 |
| Episode 1, plan v3 (the one we posted; shots 4 and 5 fell back to stills) | 328 charged, 128 refunded, net 200 credits | run 2026-10-23 | qa-ep1-v3.md:12 |
| Episode 2 | 400 credits | run 2026-10-26 | qa-ep2-v3.md:12 |
| Episode 3 | 336 credits | run 2026-10-26 | qa-ep3-v3.md:15 |
| Episode 4 | 344 credits | run 2026-10-27 | qa-ep4-v3.md:16 |
| Episode 5 (quoted 368; two stills refunded) | 232 credits charged | run 2026-10-27 | qa-ep5-v4.md:16 |
| Provider cost we measured for the posted episode 1 | about 0.27 USD | run 2026-10-23 | qa-ep1-v3.md:14 |
| Provider cost we measured for episode 2 | 0.5449 USD | run 2026-10-26 | qa-ep2-v3.md:14 |
| Provider cost we measured for episode 3 | 0.4485 USD | run 2026-10-26 | qa-ep3-v3.md:17 |
Total credits (all eight episode renders: 290 + 328 + 328 + 200 + 400 + 336 + 344 + 232 (character sheets not included)): 2,458 credits, which is $49.16 at the $0.02-per-credit list rate. Plan prices differ; see what things cost.
Credits are what the product charged; refunds for failed clips are included. The USD rows are our own measured provider cost, not an invoice and not the customer price. Most of the first-day spend went on episode 1 because we kept finding defects before posting.
What the platforms showed
| What | Number | As of | Source file : line |
|---|---|---|---|
| YouTube, episode 1: views, and share who swiped away | 112 views, 7.4% stayed to watch, 92.6% swiped away, average view 0:19 of 0:32 | 2026-10-26 08:20-08:45 ICT, 12.2 h after posting | measurements-2026-10-26-evening-posts.md:21-26 |
| YouTube, episode 1: shape of the views | about 35 views by 0.5 h, about 110 by 1.4 h, then flat; 9 engaged views of 112; 0 subscribers gained | 2026-10-26 08:20-08:45 ICT, 12.2 h after posting | measurements-2026-10-26-evening-posts.md:21-28 |
| Instagram, episode 1: views | 15 views, all from non-followers | 2026-10-26 08:20-08:45 ICT, 12.2 h after posting | measurements-2026-10-26-evening-posts.md:12 |
| X, episode 1: views on the post page | 3 views | 2026-10-26 08:20-08:45 ICT, 12.0 h after posting | measurements-2026-10-26-evening-posts.md:14 |
| YouTube, episode 2: views, share who swiped away | 6 views, 75.0% swiped away | 2026-10-28, read 06:40-09:40 ICT (about 46-49 h after posting) | measurements-2026-10-28-autoshort.md:8 |
| YouTube, episode 3: views, share who swiped away | 8 views, 85.7% swiped away | 2026-10-28, read 06:40-09:40 ICT (about 22-25 h after posting) | measurements-2026-10-28-autoshort.md:10 |
| Instagram, episodes 2 and 3: views on the reels grid | 15 and 46 views | 2026-10-28, read 06:40-09:40 ICT | measurements-2026-10-28-autoshort.md:23 |
| Facebook Page, episode 3 | reach 61, 66 views, average watch 6 s | 2026-10-28, read 06:40-09:40 ICT | measurements-2026-10-28-autoshort.md:26 |
| TikTok, posts of 26-27 Oct | 0 views | 2026-10-28, read 06:40-09:40 ICT | measurements-2026-10-28-autoshort.md:20 |
| A cat-fact Short on another channel the same night (for comparison) | 111 views, 90.2% swiped away | 2026-10-26 08:20-08:45 ICT | measurements-2026-10-26-evening-posts.md:50 |
Own views are not excluded (verification plays, comment visits). The YouTube views of episode 1 came almost entirely from the Shorts feed on mobile phones, so most were not from our desktop. Small numbers: signal, not proof.
What we learned
- Episode 1 looked like a win next to our other posts that night, and it was not one: a plain cat-fact Short on another channel got about the same views with a similar swipe-away rate. YouTube seems to hand out one test batch of roughly a hundred views and stop. We call that a test burst, not an audience.
- About nine in ten viewers swiped away within the first seconds. The story was never the problem, the first second was.
- Quality gates paid for themselves: a logo on a prop and a phone copied into several shots were caught before posting, and each re-render is a paid render.
- The heroine's face still drifts between shots. We did not fix that, and viewers can see it.
- Later episodes on the same channel got single or double digit views. A series did not build an audience in the days we measured.
What we would do differently
- Say the hook in the first second, spoken and as a caption, with no establishing shot.
- Run the same episode with two different first two seconds and compare the swipe-away share instead of raw views.
- Fix the character sheet (no props in the hand) before the first render, not after the fourth.
The public videos
Links as recorded in our publishing ledger. Platforms can remove or restrict a video later.
Try the same kind of prompt
Plan a fixed-character series in a prompt before you spend credits: the draft and the cost estimate are free, and a render waits for your yes.
Open the prompt libraryConnect AutoShort to Claude, ChatGPT, Cursor or Gemini CLIFree script preview
A prompt does not make a video perform. It makes the script and the cost visible before you spend credits. Results of ours are above.
Data sources
Files in our internal research folders that every number on this page comes from:
- growth-ops/research/drama-remix-2026-10-23/README.md, qa-ep1.md, qa-ep1-v2.md, qa-ep1-v3.md, qa-ep2-v3.md, qa-ep3-v3.md, qa-ep4-v3.md, qa-ep5-v4.md
- growth-ops/research/animal-comedy-series-2026-10-23/measurements-2026-10-26-evening-posts.md
- growth-ops/research/measurements-2026-10-28-autoshort.md
- growth-ops/ledger/published.md (public links)