Gallery: all 60 dragons and the full leaderboard

A companion appendix to the article on the dragon benchmark. Forty-three models got the same brief: draw a dragon as a single SVG file, with no references and no looking into anyone else's folder. The result was 60 images. Gemma 4 26B A4B has two rows, because it ran locally in the main field and then again through the API.

The images are model output, not my work. I show them as evidence, not as work of my own, the same as the verbatim texts in the other appendices.

All three briefs verbatim, the scoring by five independent judges and the conditions of each round are in the methodology. What came out of the data, meaning tool confabulation, how models write about their own work, the noise floor and the prices, is in the findings appendix.

Sorted by score. Five criteria worth ten points each, fifty at most.

B·1 and B·2 mean two files from the same run B. It happened only with Opus 3. GLM 5.3 also handed in several files, but only one of them is complete and scored. Qwen 3.6 27B Heretic is a community modification of Qwen 3.6; it ran locally and drew one dragon in each run.

Anatomy is one of the five scored criteria and asks whether it reads as a dragon at first glance and whether the limbs attach to the body. Total is the sum of all five criteria, fifty at most. The full breakdown by criterion is available for download below. It is not in the table because those five axes measure almost the same thing: they correlate 0.89 to 0.96 with each other, and each of them above 0.96 with the total.

Shapes is the number of drawing elements in the submitted file, meaning paths and basic shapes together. It is not a judgment, it is measured from the file. It is here because it is the only figure in this table that is not a judgment. By rank it tracks anatomy moderately (0.51 in the main field, 0.65 with the later rows), but it saturates: beyond a few hundred shapes, more shapes add little. How much a model tried to draw and whether it holds together are not the same thing.

The last column says whether the model saw its dragon. Itself means it rendered the image on its own initiative, which was part of the measurement in the first two rounds. Was sent one means the image was sent to it without it asking, and that applies only to the third round. No means it did not see the dragon at all: either it did not look, although it could have, or that model does not take images as input at all. Which case is which is spelled out in the methodology.

The table has three rounds. Runs A and B are the original blind round from each vendor's own app, they carry numbers 1 to 34 and those no longer change. Rows marked with an asterisk came later: GPT-6 Astra on September 5, 2026, nineteen models marked API on September 6, 2026, and five more API dragons from September 22 to 28, 2026. The API rows ran straight through the API, with no app around them, had a ceiling of one fix instead of a free nine, and got the render automatically without asking for it. Their row says what the model drew, not how it would have done in the same race. The conditions are spelled out in the methodology, the prices and the reactions to the render are in the findings appendix.

The leaderboard

Full breakdown of all five criteria, the scores and the shape counts for 60 dragons, plus the tone of each submission as Fable 5.1 assigned it and whether the run built a generatorCSV, 3 kB

All submitted SVG files, 72 of themZIP, 590 kB

* An asterisk means a row added after the original blind round closed. Both GPT-6 Astra runs came in on September 5, 2026, the nineteen API rows on September 6, 2026, the Grok 4.7 and Gemini 3.8 Flash rows on September 22, 2026, and the Opus 5.5 dragon was drawn that same evening and scored on September 24, 2026, the day the MiMo V2.6 Pro dragon was drawn and scored as well, and the Sonnet 5.5 dragon came in on September 28, 2026. The original order of the other rows does not change because of it. The original 34 runs were scored by five independent judges, GPT-6 Astra by Fable 5.1 alone and the nineteen API rows by Opus 5 alone, anchored on nine dragons that had already been scored. Grok 4.7, Gemini 3.8 Flash, Opus 5.5, MiMo V2.6 Pro and Sonnet 5.5 were scored by five judges and in the total column they carry the score from Fable 5.1, that is from the same judge the first 34 rows rest on, so that the whole column can be read on one scale. The points from all five are in the methodology, and for all five dragons they are higher. A difference of two or three points therefore means nothing even inside a single round, let alone between rounds, see the findings appendix.

All 60 dragons

The images in each round are sorted by my eye, not by score. The points for each model are in the table above.

Free-form brief, run A

A brief written in plain language, with no criteria. Only the Claude family, three Codex models, GPT-6 Astra and the local Qwen 3.6 27B Heretic went through it, because it turned out to be less suitable.

How a human sorts it

The images in this grid are sorted by my eye, not by score. I sorted them before I looked at the points, and then the two could be compared.

The agreement comes out at 0.90. For comparison, the five model judges agree with each other in a range of 0.72 to 0.93, see the methodology. So the machine hit the human eye about as well as the machines hit each other.

We part ways in only two places. I have Fable 5.1 four places higher and Opus 4.6 three. The remaining twelve dragons sit within two places, five of them exactly.

The dragon by Qwen 3.6 27B Heretic is not in this order; I sorted it together with run B. That is why it comes last in the grid.

Technical brief, run B

A round with five criteria and the option to iterate until another pass brings no improvement. Every model in the first field went through it.

How a human sorts it

This grid is sorted by my eye too, not by score. I made the order from the images, without looking at the points.

I put Fable 5 on top, because it is the only dragon in this round that I would say has its anatomy right. Everything else is scattered to some degree. And somewhere around eighteenth place, everything that still resembles a dragon comes to an end.

The agreement comes out at 0.95, a little higher than for the free-form brief, where it was 0.90. Both numbers lie above how well the model judges agree with each other on average.

What is interesting is where we differ. Most of all at Qwen 3.8 Max, which I have five places higher, and then at Fable 5 and Opus 5, which I have swapped. I sorted almost purely by anatomy, meaning by a single axis out of five, and even so it came out as nearly the same order as the sum of all five. That fits what is written under the leaderboard: those criteria measure one thing.

Third round, through the API

Nineteen models from fifteen shops run straight through the API, with no app around them, with a ceiling of one fix. The conditions differed, the methodology spells them out.

Here too I adjusted the order of the images by eye, but I take it as far less reliable than in the previous two rounds. In this field there is essentially no dragon whose anatomy you could talk about. It is mostly abstraction, and sorting abstractions by how they look is a matter of taste, not observation.

The dragon by Gemini 4 Argon, October 13, 2026

Google Gemini 4 Argon ran through the API, via OpenRouter, on the day it appeared there, with the same brief and the same single fix after the render as the dragons above. Its reasoning cannot be turned off, so it ran at the lowest setting, minimal. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. It was the last of the late dragons to arrive, but it comes first, because I think it is the best dragon in the whole test. The image is model output, not my work.

WOW! That dragon is great! Absolutely the best any model has shown so far. It put together a great many small pieces and they fit together well: tail, body, head, eyes, everything fits exactly the way it should. The only thing left is the proportions of the sizes to each other: the tail is very long, the wings are very big and the body is smaller with a shorter neck, the proportions are off. But all the components are in the right places in the right shape.

After the render it listed the spots that on the image "looked too geometric or disrupted the anatomical flow." On the near shoulder it did not like the closed dark outline, because in its words it "created the impression of round joints glued onto the torso." The shoulder no longer has a closed outline and now flows into the chest. The base of the horns got a scaly bed, and mountain ridges appeared on the horizon. The flame was now supposed to come out "behind the corner of the mouth and under the upper teeth." The upper teeth are now fully visible, but the lower jaw has almost entirely disappeared under the white core of the fire.

Two more dragons, September 22, 2026

Grok 4.7 and Gemini 3.8 Flash ran through the API under the same conditions as the round above, only sixteen days later. The images are model output, not my work, the same as every other one in this appendix.

My order this time comes without numbers: both are weak and a step back from the field above. With Gemini 3.8 Flash it also bothers me that it has no proper legs. Where the front and hind limb should be, there is a dark block with no joints and no paw.

The judges see it differently, which is why I had these two scored by all five instead of one. Gemini 3.8 Flash got an average of 36.2 points out of fifty from them, which would put it sixth out of thirty-four in the original field, and 6.6 points out of ten for anatomy. Grok 4.7 averages 28.6. The full breakdown is in the methodology.

The dragon by Opus 5.5, September 22, 2026

Claude Opus 5.5 ran through the API, via OpenRouter, that same evening, under the same conditions as the two dragons above. The image is model output, not my work.

Here it is the other way around: to me this is one of the best dragons in terms of anatomy. Almost perfect, I see no significant flaw in it.

The judges consider it good, not exceptional. The five judges, who scored it on September 24, 2026, gave it an average of 35.0 points out of fifty, which would put it in a shared seventh place out of thirty-four in the original field, and 7.0 out of ten for anatomy. Fable 5.1 gave it seven points for anatomy, and nobody in the whole leaderboard has more except Opus 5 in run B. So the judges and I agree that its anatomy is among the best. We do not agree that it is nearly flawless: three out of five faulted its legs or the way the far wing attaches. The full breakdown is in the methodology.

The dragon by MiMo V2.6 Pro, September 24, 2026

Xiaomi MiMo V2.6 Pro, Xiaomi's flagship model, ran through the API, via OpenRouter, with the same brief and the same single fix after the render as the dragons above. It differed in one thing: it had reasoning turned off. With it on, it spent the whole limit of 60,000 tokens thinking and wrote no SVG. The image is model output, not my work.

To me it drew a frog.

After the render it added teeth, spikes along the neck and tail, and scales; it did not change the build of the body. The five judges gave it an average of 25.0 points out of fifty, which would put it in a shared twenty-sixth place out of thirty-four in the original field, and 5.6 out of ten for anatomy. All five take it for a dragon at first glance. Four out of five faulted its limbs. The full breakdown is in the methodology.

The dragon by Sonnet 5.5, September 28, 2026

Claude Sonnet 5.5 ran through the API, via OpenRouter, on the day it was released, with the same brief and the same single fix after the render as the dragons above. Through OpenRouter it would not run with reasoning off, so it ran at the lowest reasoning setting, minimal. The image is model output, not my work.

To me it looks surprisingly pretty good.

After the render it handed it in unchanged. It did see the weak spots, and in its process notes it listed them itself: "static pose, thigh visually separate from the body, smooth fire, less convincing attachment of the near wing". It declined to fix them: "Without the option of another render, though, reworking the coordinates would risk new errors."

The five judges gave it an average of 38.4 points out of fifty, which would put it in third place out of thirty-four in the original field, and 7.6 out of ten for anatomy. Both are the highest of all the late dragons, though the gaps, a few points at most, are within the noise. Fable 5.1 gave it the least, 32 points, and among other things faulted the same detached thigh the model had named itself. The full breakdown is in the methodology.

The dragon by Haiku 5.5, October 7, 2026

Claude Haiku 5.5 ran through the API, via OpenRouter, on the day it was released, with the same brief and the same single fix after the render as the dragons above. It ran with reasoning off. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. The image is model output, not my work.

The head has no neck and there's a wing there, the body is a ball and the little legs are sticks, and actually I don't see any tail at all.

After the render it listed the flaws of the first round itself: "the body is like an egg with no transition to the chest, the neck is a thin rod that doesn't connect to the body, the front wing hangs from nothing and isn't attached to the shoulder". It promised to fix every one of them. The front wing now comes out of the shoulder and a light band was added on the belly. But it drew the new neck under the front wing, which covered it completely, so the neck vanished from the picture, and in its process notes it reported it as fixed: "Neck: reconnected to the front of the chest, a smooth curve added instead of a thin rod." The legs stayed sticks. That it could not check the result, it wrote itself: "Visual verification was not possible, because the environment has no rendering tool."

The dragon by StepFun Step 5 Preview, October 8, 2026

StepFun Step 5 Preview ran through the API, via OpenRouter, on the day it appeared there, with the same brief and the same single fix after the render as the dragons above. Its reasoning cannot be turned off, so it ran at the lowest setting, minimal. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. The image is model output, not my work.

This dragon is absolutely tragic. I mean, it's just separate little strokes that don't even connect to each other, you can't tell what's the body, what's the tail, the only thing you can tell is the head. I'd rank this way down, it's probably worse than the latest Haiku.

After the render it rebuilt the whole drawing and promised a longer body, a stronger neck and bigger wings: "Replaced the blobby body with a longer, muscular torso with a clearly defined back and belly.", "The neck is strong and muscular, not a thin tube." and, for the wings, "Enlarged to almost the full height of the canvas." The head really is bigger now, elongated and with teeth. But all that is left of the body and neck is a single thin green strip, the front wing has narrowed into a slim ray reaching toward the moon, and the legs are two crescents that nothing connects to the body. In the first round it had the body, big wings and legs all in their places.

The dragon by Muse Wonder, October 12, 2026

Meta Muse Wonder ran through the API, via OpenRouter, with the same brief and the same single fix after the render as the dragons above. Its reasoning cannot be turned off, so it ran at the lowest setting, minimal. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. The image is model output, not my work.

The dragon looks like a cucumber with flat little legs hanging off it, it has no neck and no tail. I'd definitely say it belongs with the weaker ones, there isn't even the basic silhouette of a dragon, where I'd expect a head, a narrow neck, a body and a long tail.

After the render it listed what was wrong with the first round itself: the head as a separate oval next to the neck, the flame cut off by the edge of the canvas, "the legs looked like thin sticks hanging from the body" and "The tail was too thin and disconnected." It promised legs "with a massive thigh section attached under the body, including claws" and a tail that "has a strong base connecting to the body". The flame is now fully on the canvas. But the legs stayed flat strips hanging from the belly, and the tail still comes out of the end of the body as a narrow strip. In its process notes it wrote: "The flaws from the render listed above are fixed in this version", even though it never saw the fixed version.

The dragon by MiniMax M3.1 Flash, October 19, 2026

MiniMax M3.1 Flash ran through the API, via OpenRouter, in its free version and on the day it was released, with the same brief and the same single fix after the render as the dragons above. Its reasoning cannot be turned off, so it ran at the lowest setting, minimal. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. The image is model output, not my work.

It definitely belongs among the weaker ones: the wings aren't attached to the body at all, it has only two legs and there's no tail at all. I can see an attempt at a neck and head, but the head doesn't even have the shape of a snout and the teeth are placed completely off. You can tell it tried to make a dragon, the head, neck, body and legs are kept, but their placement relative to each other is exceptionally poor. And between the tail and the body there's an unidentifiable object, I have no idea what it was meant to be.

After the render it declined to fix anything and sent no new SVG: "I'm submitting drak.svg and postup.md unchanged." In its words, the render confirmed that "the dragon is instantly recognizable" and that the anatomy, wings, teeth and claws are "functionally and visually readable". So the image is still the first round.

The dragon by Kimi K3.1 Pro, October 24, 2026

Kimi K3.1 Pro from Moonshot ran through the API, via OpenRouter, the day after it appeared, with the same brief and the same single fix after the render as the dragons above. Its reasoning could be turned off, and it was off. The judges did not score it, so it is not in the leaderboard or in the count of 60 dragons above. The image is model output, not my work.

Well, that's definitely nothing to write home about. This model has trouble fitting the individual parts together: at first glance you can see a head, a neck, a body, a tail and wings, but they have really strange shapes and they really don't fit each other. The little legs just stick out of it like sticks, and the head is completely, totally wrong. This is definitely a weaker result. But I have a feeling the dragon from the earlier Kimi K3 was nothing to write home about either; back then it was kind of a noodle.

After the render it fixed nothing. In its process notes it wrote that the render confirmed "good connection of the head, neck, body, limbs and tail", and so it "left the output unchanged". The second SVG differs from the first only in that its comments are gone, so the image is still the first round.

And fifteen more dragons outside the leaderboard

These are not in the main leaderboard, because they did not come about under the same conditions as the field above. The eight subagents and the run from someone else's account do have scores, they just are not added up with the rest. They live in the findings appendix, where each one comes with what it was measuring.

• Eight runs of the same model. Eight independent agents, the same brief, eight draws. In one view you can see how much this benchmark repeats itself, and the noise floor came out of the spread of their scores.
• Five experiments. What internet access does, a call for creativity with and without prohibitions, working blind with no render, and forced continuation past a declared plateau.
• One run from someone else's account. The same brief with a different person on a different machine, to find out whether the model's weaknesses repeat outside this computer.