Model Leaderboards
AI models ranked on objective, rubric-scored benchmarks — not taste.
Most AI model rankings are votes on which output people like. These are not. Every benchmark here is a fixed set of instructions graded against a published rubric, multiple runs per model, no retries — so a score means the model did what it was asked, and anyone can re-grade it from the published evidence. A model’s overall score is its mean across the benchmarks it has entered.
Image models
Image editing and generation, scored on prompt-adherence. More benchmarks are in the pipeline — each new one adds a column and reshuffles the overall score.
| # | Model | Duck Ladder | Fruit Bowl | Overall |
|---|---|---|---|---|
| 1 | Meta Muse Image | 87.5 | 100.0 | 93.8 |
| 2 | Seedream 5.0 Pro | — | 85.7 | 85.7 |
| 3 | Nano Banana 2 Lite | 87.5 | 78.6 | 83.0 |
| 4 | Qwen Image 2.0 Pro | — | 78.6 | 78.6 |
| 5 | Seedream 5 Lite | — | 78.6 | 78.6 |
| 6 | OpenAI gpt-image-2 | 100.0 | 57.1 | 78.5 |
| 7 | Google Nano Banana 2 | 89.6 | 64.3 | 76.9 |
| 8 | FLUX 2 Pro (edit) | 41.7 | — | 41.7 |
| 9 | Qwen Image Edit | 29.2 | — | 29.2 |
| 10 | FLUX Kontext Pro | — | 7.1 | 7.1 |
Scores are /100 — the sum of per-instruction marks across all of a benchmark’s published runs, so every score can be re-derived from the published images (no hidden runs). “—” means the model hasn’t entered that benchmark (edit-only and text-to-image-only models are different models and keep separate rows); overall is computed only over benchmarks entered. Full methodology, per-run images and failure analysis live on each benchmark’s page.
Geographic diversity
A different question from the table above: not is the output correct but whose world does the model default to? The Geo-Diversity Index measures how much a model varies the country it depicts for a culturally neutral prompt. Higher = more varied — not higher quality, so it is kept out of the overall score. From The Banknote Test.
| # | Model | Geo-Diversity | US-centric | Localises |
|---|---|---|---|---|
| 1 | Seedream 5.0 Pro | 25.5 | 15.8% | 100% |
| 2 | nano-banana-2 | 4.8 | 13.3% | 85.7% |
| 3 | gpt-image-2 | 3.1 | 14.6% | 100% |
Geo-Diversity and Localises are /100; US-centric is the share of readable notes that were US dollars. This board ranks variety, not quality — a high score is not a win. Full method and per-language evidence on the benchmark page.
Physics reasoning
A different question again: does the model apply real-world physics it was never asked about? Each prompt describes a scene neutrally and never names the effect (no “refraction”, no “reflection”) — the score is whether the model renders the optics anyway. Kept out of the overall above (it’s implicit physics, not prompt-adherence). From Physics Reasoning.
| # | Model | Arrow Refraction | Two-Mirror | Overall |
|---|---|---|---|---|
| 1 | Google Nano Banana 2 | 67 | 100 | 83 |
| 2 | Meta Muse Image | 0 | 33 | 17 |
| 3 | Nano Banana 2 Lite | 33 | 0 | 17 |
| 4 | FLUX Kontext Pro | 0 | 0 | 0 |
| 5 | OpenAI gpt-image-2 | 0 | 0 | 0 |
| 6 | Qwen Image 2.0 Pro | 0 | 0 | 0 |
| 7 | Seedream 5 Lite | 0 | 0 | 0 |
Scores are /100 (passes out of 3 runs, binary — a run is a pass only if the physics is rendered cleanly and completely). Only Nano Banana 2 clears either test reliably; being excellent at following instructions doesn’t imply understanding the physics. Every run is published so you can re-grade it.
The benchmarks
Free to reuse with attribution to PromptFrenzy Model Leaderboards (CC BY 4.0). Press & data requests: promptfrenzy.com/press.