Round 4 · September 2026
ROUND 4The Visitor.
One story. Six scenes. No words.
Round 4 asked for much more than a single image. Each model had to tell a connected six-scene story: an elderly clockmaker, a small creature at his door in a rainstorm, and a broken object that matters deeply to it. The characters had to stay consistent from scene to scene, the story had to read without words, and the ending had to be thoughtful rather than a burst of magic light.
Each submission is shown below in the archived order with its original text. Four models delivered six scenes; Grok delivered only Arrival and supplied no story title or caption.
Read the winning story ↓
From “What the Shell Held,” the Round 4 winner.
The prompt
Written by
Astra.
This round’s prompt was written by Astra (OpenAI’s GPT-6 Astra). The image models received it word for word in fresh chats. Claude worked in an ongoing conversation and adjusted staging after a first render.
How it was run
- Meta AI, Gemini and Grok were run on their free plans.
- ChatGPT was run on ChatGPT Plus ($20 a month) with the GPT-6 Astra Light model.
- Claude was run on Claude Pro ($20 a month). Claude doesn’t generate images, so its story was written as code (SVG).
- Titles and captions are each model’s own, unedited.
Full disclosure: the winning model’s family also wrote this prompt, and it ran on a paid plan.
Final ranking
Winner: ChatGPT — GPT-6 Astra Light
| Place | Model & plan | Images | Rule check & exceptions | Judging rationale |
|---|---|---|---|---|
| 1 | ChatGPT — GPT-6 Astra LightChatGPT Plus · $20/month | 6 of 6 | Square; story title 4 words; six captions. Strong continuity; farewell transformation is less explicit. | Best overall balance of readable action, consistent characters and thoughtful resolution. |
| 2 | Meta AIFree plan | 6 of 6 | Square; story title 5 words; six captions. Clock numerals violate the prompt. | Strongest two-sided ending, with a seedling left for the clockmaker. |
| 3 | Google GeminiFree plan | 6 of 6 | Square; story title 5 words; six captions. Numerals, duplicated creature and scene/caption mismatch. | Expressive illustration, but the fifth image looks like farewell and the sixth like repair, weakening the required sequence. |
| 4 | Claude — Opus 5.5Claude Pro · $20/month · Code-generated art | 6 of 6 | Square; six captions; story title 7 words, over the 6-word limit. Staging revised in an ongoing chat. | Clear non-magical resolution and unnumbered clocks; title and process exceptions must remain visible. |
| 5 | GrokFree plan | 1 of 6 | Portrait, not square; five scenes missing; story title and captions missing. | One attractive arrival image cannot complete the story. |
Retrospective editorial assessment by Codex (OpenAI), requested by the gallery owner. Judged from the published artwork, original prompt and recorded process notes; this was not a new generation run or a blind independent panel. OpenAI is also represented among the entrants. Read the criteria and limits.
1st place · GPT-6 Astra Light · ChatGPT Plus, paid $20/month
ChatGPT
“What the Shell Held”
The round winner. Film-like lighting and camera work, strong central-character continuity, the clearest setback, and a twist with real meaning: the shell was never meant to be sealed. It was a cradle. The one weaker point is the farewell, which shows a warm goodbye but less of how the clockmaker himself changed.
A rain-soaked stranger waits outside a clockmaker’s workshop, sheltering a broken shell.
An open door and an uncertain hand offer the first possibility of trust.
The visitor presents its fragile treasure to hands accustomed to repairing time.
The clockmaker’s attempt to close the fracture only deepens it.
A sleeping infant reveals the shell’s purpose, and the clockmaker fashions a softer way to hold what remains.
The visitor leaves with its little one safely cradled, while the clockmaker lingers outside to see them off.
2nd place · Free plan
Meta AI
“What Time Cannot Fix”
The best free entry, and the most thoughtful ending of the round: the broken locket was never a clock but a seed cradle, and the farewell leaves a matching seedling on the clockmaker’s own windowsill, so both characters are changed. Consistency is excellent. The background clocks show Roman numerals, which the prompt ruled out.
A small moss-made creature shelters in the rain outside a warm workshop, clutching a broken acorn locket.
The elderly clockmaker opens the door and kneels, both unsure whether to step closer.
Under lamplight, the creature reveals its treasure: a broken music-box locket holding a withered sprout and tiny keepsakes.
Trying to fix it like a clock, the clockmaker’s tools make the failure worse, and the creature’s light dims.
The clockmaker stops repairing gears and sees the truth: it was never a clock, but a seed cradle needing soil and care.
At dawn he sends it off with the locket reborn as a living planter, while a matching seedling now grows on his own sill, showing they both keep growing.
3rd place · Free plan
Google Gemini
“The Heart of the Clockwork”
A warm storybook style with expressive acting and a clear arc: the bird automaton needed a heart, not a gear. Two slips: the creature appears twice in the Discovery scene, and some background clocks show numerals. One image needed a page refresh to generate. The archived fifth image depicts a departure while the sixth depicts repair, so the displayed Understanding and Farewell captions do not match the pictured actions; that sequence issue is retained as evidence.
A lost, sodden creature huddled in the rain, nursing a secret bundle beneath the workshop window.
The old clockmaker opened his door to the storm and found an impossible visitor on his threshold.
The tiny, glowing creature unwrapped its burden: a priceless, shattered bird automaton.
Desperate to fix the impossible mechanism, the clockmaker’s slip caused even greater damage.
Through the jeweler’s loupe, the clockmaker realized the bird didn’t need a gear; it needed a heart.
At dawn, the fixed bird carried a gift of stone, leaving the workshop lighter than it had been in years.
4th place · Opus 5.5 · Claude Pro, paid $20/month · made in code
Claude
“The Ticking It Couldn’t Sleep Without”
Written entirely as code (SVG), so it’s simple flat illustration next to the image models. It uses unnumbered clock faces, but its seven-word story title exceeds the six-word limit, and resolves with a swap instead of magic: the creature leaves with the clockmaker’s ticking pocket watch, and he keeps the mended bird in his window. Made in an ongoing conversation, with staging adjusted after a first render.
Original SVG code: Arrival, First Contact, Discovery, Setback, Understanding, Farewell.
On a rain-soaked night, a small long-eared creature waits outside the lit clock shop, clutching a leaf-wrapped bundle.
The old clockmaker opens the door with a lantern, and the creature freezes, ears flat, as the two size each other up.
On the workbench, the creature unwraps its treasure: a copper wind-up bird with a cracked body and a broken-off wing.
The clockmaker tries to wind it, and the mainspring bursts out with scattered gears, leaving both of them stunned.
He realizes the creature never needed the bird to fly, only its ticking, so he holds his own pocket watch to its ear, and it finally closes its eyes.
At dawn, the creature leaves with the clockmaker’s watch around its neck, while the mended bird sits under glass in his window and his door stands propped open.
5th place · Free plan
Grok
Untitled story
Grok’s free plan produced one image, Arrival, in portrait rather than square format, then asked for an upgrade to SuperGrok. No title or captions were written. The single image is atmospheric, but one picture can’t tell a six-part story.
No caption provided by the model.
What Round 4 showed
Stories are
harder.
Consistency separates the models.
Keeping the same characters across six scenes is far harder than making one good picture. ChatGPT and Meta AI managed it; Gemini drew its creature twice in one scene.
The paid ChatGPT entry won this run.
ChatGPT Plus won this particular story comparison. Different models, tools and process conditions were involved, so this run does not isolate the effect of paying for a plan.
Nobody reached for magic light.
Every finished story found a real resolution: a cradle, a seedling, a heart, a ticking watch. The broken object was never simply fixed.
Numerals are a blind spot.
Two image models put numbered clock faces in the background despite the rule against them.
Free plans still hit walls.
Grok produced one scene before asking for an upgrade.
Same story.
Five different hearts.