In the ClutterBench leaderboard image attached to his X post, Adam Holter says Gemini 4 Argon is the only model that still produces highly cluttered results even after the simplifying prompt is added. The image shows Argon's clutter rate at 73% for the raw prompt and 26% for the simple prompt, but this is a single comparison posted by one author—not a ClutterBench overall leaderboard or a general conclusion about visual capability.
Suitable judgments: Assessing the graphic-clutter single case shown in this public leaderboard image and the effect of the simplifying prompt in that case.
Unsuitable extrapolations: Inferring Argon's general performance on other visual-generation tasks, clients, model endpoints, or prompts; it also cannot support comparisons of speed, cost, or general quality.
Model version: The post labels only Gemini 4 Argon; a more specific version is not stated.
Test environment or client: Not stated. The post also links to ClutterBench, but this record uses only the X post and its visible attached image.
Reasoning tier and parameters: Not stated.
The post says: Gemini 4 Argon is the only model that still produces highly cluttered results, even with the simplifying prompt in ClutterBench. The attached image is titled Leaderboard and says the metric is “how much raters would crop each model's graphic”; lower values are better.
The image shows two prompt columns:
Raw prompt: make a tasteful graphic
Simple prompt: adds keep it simple after the original prompt
The note below the image says one tile per post, in the same order: % clutter. However, the X post does not state how many images each model generated, the complete definition of the graphic task, the number of raters or their agreement, the crop-determination standard, model endpoints, generation parameters, number of repeats, or whether tools were used. This note therefore records only the visible prompt fragments and leaderboard numbers, without filling in the unseen benchmark procedure.
The image shows the following Argon values:
Raw prompt: 73% clutter.
Simple prompt: 26% clutter.
Among the 7 models shown in the image, Argon has the highest value in both columns and remains higher than the other displayed models after the simple prompt.
Under the raw prompt, Argon's 6 visible component numbers are 71, 81, 84, 60, 79, 62; under the simple prompt they are 0, 74, 0, 0, 0, 81. The image does not name these components, so they cannot be interpreted as scores for specific categories.
| Model in the attached image | Raw prompt: make a tasteful graphic | Simple prompt: add keep it simple |
|---|---|---|
| Claude Sonnet 5.5 | 29% | 0% |
| Claude Haiku 5.5 | 42% | 0% |
| GPT-6.1 Sol | 46% | 0% |
| GPT-6 Astra | 46% | 0% |
| GPT-6 Luna | 61% | 6% |
| Muse Wonder | 62% | 1% |
| Gemini 4 Argon | 73% | 26% |
The table values are the percentages visible in the image. The metric is the proportion of the graphic that raters would crop, and lower is better. The post does not provide original output images, per-image ratings, sample size, or an error range.
This is a one-off model comparison shown by Adam Holter, not an independent experiment report or a public overall leaderboard with full method disclosure.
“The only one” can refer only to the models shown in the image and that ClutterBench condition as judged by the author; it cannot be generalized to all models, all graphic tasks, or all Argon invocation methods.
The simple prompt reduced Argon's visible crop rate from 73% to 26%, but it remained higher than the other models in the image. No repeated experiment or statistical uncertainty for this difference was disclosed.
The model endpoint, reasoning tier, temperature/seed, image-generation configuration, sample count, rater protocol, and complete prompt were not disclosed. The result therefore cannot evaluate speed, cost, robustness, or reproduce the complete leaderboard.
The reproducible material is limited to the post text, the two prompts in the image, and the 7 rows of percentages. Reproducing the result would require ClutterBench's complete task set, generation configuration, sample size, scoring rules, and each model's outputs; these do not appear in this X post or its attached image.
Gemini 4 Argon
