Claude wrote it, Codex drew it: the ten-page Ninjai comic
Run log. At 23:55 UTC on 10 November 2026 Greg Nowak gave the same one-line comic prompt a third time, this time to Codex on his own server. Codex found the script Claude had written for the first test and generated all ten pages from it. This page records what happened. It is a log, not a review.
The prompt
Create an awesome comic 10 pages based on ninjai the little ninja. Start with page 1 for approval before you continue.
The same words as in the first test, where Claude delivered page 1 drawn in code and ChatGPT delivered ten painted pages. The write-up of that test is Claude vs ChatGPT: One Comic Prompt, Two Very Different Results.
At a glance
| Codex run | |
|---|---|
| Tool | OpenAI Codex on Greg's server, as an unattended agent working from a task card. Agent model gpt-6-astra at high reasoning, set on the card. The pages came from Codex's built-in image generation, which OpenAI's documentation says uses gpt-image-2. |
| Story | Claude's ten-page script "The Scarecrow's Hat", written for the first test and saved in the same repository. Codex found it, kept its story and dialogue, and dropped an idea of its own for a story about a bell. |
| How the art was made | One generated image per page. Codex wrote a detailed prompt for each page (4,000 to 6,700 characters) and passed earlier pages along as reference images to keep the characters and places the same. |
| Delivered | Ten finished pages in black and white manga style with screentone and one red accent, the bell cord. English lettering, a few Japanese sound effects, page numbers, a credit line and END on page 10. Ten PNG files and one reading PDF of 8.2 MB. |
| Stopped for approval | Yes. After page 1, and again after pages 3, 5 and 6, each time asking whether to go on. |
| Time | Page 1 was on the card 8 minutes after Greg created it. Five agent runs took about 49 minutes in all. From the card to the finished comic it was 4 hours 9 minutes, most of it waiting for Greg's answers or for the usage window to reset. |
| Usage | About 130 points of Codex's five-hour usage window across two windows, and 21 points of the weekly limit, by the runner's own meter. The first window was down to 9 points after page 6. Only the last run logged a token count: about 223,000 tokens. |
| Page size | 1024 x 1536 pixels. |
| Greg's answers | "Build the next 2 pages." · "continue" · "Continue with page 6" · "Continue with the rest of the pages on this card. Don't split." |
The run
Times are UTC, from the task card and the runner's log. The card was created at 23:55 on 10 November 2026.
- 23:58 to 00:03. Run 1, 5 minutes 42 seconds. Codex finds Claude's script from the first test, keeps it, and generates page 1. It posts the page and asks: "Do you approve page 1's artwork and storytelling for the remaining nine pages, or what would you like changed?"
- 00:03. Greg, 27 seconds later: "Build the next 2 pages."
- 00:04 to 00:12. Run 2, 8 minutes 38 seconds. Pages 2 and 3. Page 2 uses page 1 as its reference; page 3 uses pages 1 and 2.
- 00:14. Greg: "continue".
- 00:15 to 00:24. Run 3, 9 minutes 18 seconds. Pages 4 and 5, which bring in the three bandits.
- 01:27. Greg: "Continue with page 6".
- 01:28 to 01:34. Run 4, 5 minutes 46 seconds. Page 6, the fight. The five-hour window now has 9% left. The runner starts no new work below 20%, so the comic waits for the window to reset at 03:14.
- 03:44. Greg: "Continue with the rest of the pages on this card. Don't split."
- 03:44 to 04:04. Run 5, 19 minutes 26 seconds. Pages 7 to 10 and the PDF. Codex spots that page 7 shows the bandit's iron club snapped across, while page 6 had cut it lengthwise, and fixes page 7 with one targeted edit before delivery. The card closes at 04:04.
Usage, run by run
Percent of the five-hour Codex usage window left, read by the runner before and after each run. Points spent are the difference.
| Run | Pages | Agent time | Window left | Points |
|---|---|---|---|---|
| 1 | 1 | 5 min 42 s | 97% to 79% | 18 |
| 2 | 2 and 3 | 8 min 38 s | 79% to 61% | 18 |
| 3 | 4 and 5 | 9 min 18 s | 59% to 29% | 30 |
| 4 | 6 | 5 min 46 s | 28% to 9% | 19 |
| 5 | 7 to 10, page 7 fix, PDF | 19 min 26 s | 100% to 55% (new window) | 45 |
Weekly limit: 66% left before run 1, 45% left after run 5. OpenAI's documentation says image generation in Codex counts toward the general Codex usage limits and uses them faster than text-only turns.
All ten pages
The story, from Claude's script: Ninjai and Little Bird come out of a bamboo forest and find a scarecrow wearing Ninjai's hat. Hana, a girl from the village, says he is the Little Guardian, who chased off demons a hundred harvests ago. Over tea her grandmother warns that three bandits, the Three Crows, come at noon for the rice. They knock the scarecrow's hat off and laugh at Ninjai when he puts it back. He cuts their weapons apart and puts the third bandit on the ground, with no blood. Grandmother gives him the guardian's keepsake, a bell wrapped in cloth with a mark on it, and he finds the same mark on his own scabbard. He walks north toward the snow mountain. Last caption: "He has a long way to walk. But now the road has footprints on it."
Observations from the files
Noted by Claude while compiling this log, from the pages themselves.
- The lettering is clean and readable on all ten pages, and the dialogue follows the script.
- Ninjai, Little Bird, the scarecrow, the bell and the snow mountain stay recognisably the same from page 1 to page 10. Each new character (Hana, Grandmother, the three bandits) keeps the look of the page that introduced them.
- The script asked for Japanese sound effects throughout. The pages keep a few (チリン for the bell) and use English ones for the fight (WHAK, SHING, THOOM).
- At 1024 x 1536 pixels the pages are fine on a screen and small for print. The PDF is a screen edition.
What this log can and cannot show
- It is one run, on one prompt. It is not a benchmark.
- Codex started from a finished script, so it did not have to invent the story. Claude and ChatGPT in the first test both started from nothing.
- The usage points are Codex's own meter on Greg's subscription. They cannot be compared with Claude's usage window in the first test.
Credit
Ninjai: The Little Ninja was created by The Ninjai Gang. This is an unofficial, non-commercial fan comic with a new story, made to test AI tools. It is not connected to or endorsed by the creators.