Methodology
The GenGrades methodology is a fixed test set per dimension, an anchored 0 to 10 scale, public weights per category, and a rule that nothing is scored by impression. Every model in a category runs the same tests, so a 7 means the same thing on every page, and every result can be reproduced from the prompts and outputs we publish.
Last updated
Each dimension has a fixed set of prompts and inputs and a rubric with written anchors. We do not score from a demo or from memory of what a model did last month.
Model capability is judged on the outputs. Product experience is judged on using the tool: onboarding, pricing, limits, export. A tool page shows both.
A category decides how much each dimension matters and publishes it. No universal champion: the home page labels what each tool is best for, not one winner for everything.
Outputs are scored with the model name hidden whenever the output itself does not give it away. A large language model may pre-screen, a person sets the final score.
A new version gets a new run and a new score. Old scores stay published, so improvements and regressions are visible.
One-sentence conclusion, who it is for, who it is not for, and the single flaw most likely to make you walk away. The not-for list is never softened.
The exact prompts and the outputs they produced are published, watermarked, so readers can check our work. This is where our credibility comes from.
Every rubric is written against these five bands, so a score is a statement about usability, not a ranking position.
| Band | Label | Meaning |
|---|---|---|
| 0 to 2 | Unusable | Fails the task outright or produces something that cannot be used without redoing it. |
| 3 to 4 | Poor | Gets part of the way; the result needs substantial correction before it is usable. |
| 5 to 6 | Adequate | Does the job for everyday use; visible flaws, but nothing that stops you. |
| 7 to 8 | Good | Reliable, with only minor flaws; what most people would be happy to ship. |
| 9 to 10 | Best in class | As good as anything available on this dimension at the time of testing. |
Nine common dimensions apply to every model. Each modality adds its own, with its own test set. Tools are also graded on product experience.
Common to every model
By modality
Product experience, for tools
The total is the weighted mean of the dimension scores, with the weights of the category the entry is graded in. Weights are published on each category page and can be changed, which recomputes every total in that category. Below is the set used for image generation as an example; the category page is authoritative.
| Dimension | Weight |
|---|---|
| Prompt adherence | 25% |
| Output quality | 25% |
| Price | 15% |
| Text rendering | 10% |
| Consistency | 10% |
| Speed | 10% |
| Licensing | 5% |
When a model ships a new version we re-run its test set and publish a new, dated score. Old scores stay. Each re-run has a cost cap.
Every price, limit or claim about a vendor shows its source and the day we checked it. A stale fact is worse than none, so the date is always visible.
No aggregate rating until real users have voted. No score without a test behind it. No grade is for sale, and affiliate links, where they exist, are marked and change nothing.
Conflicts of interest
GenGrades is built by a team that also makes AI tools, among them LayerGrab, MagicRemover and the apimodels model gateway, which we also use to run many of the tests. We disclose that relationship wherever one of those products appears, and they go through the same test sets, the same blind scoring and the same published outputs as everything else. If you think a grade is wrong, the prompts and outputs are public: show us where.
Every dimension has a fixed test set and a rubric with anchors on a 0 to 10 scale. The model runs the whole set, humans score the outputs against the anchors, blind to the model name where the output allows it, and the dimension score is the mean. A large language model may pre-screen outputs, but it never sets a final score.
Because a tool is a model plus a product. Model capability scores describe the output; product experience scores describe using it, from pricing transparency to export. A tool page shows both so you can tell a great model behind a poor product from the reverse.
The total is the weighted mean of the dimension scores, using the weights of the category the entry is graded in. Weights are published on every category page, and changing a weight recomputes every total in that category.
We re-run the test set on the new version and publish a new score with its own date and version label. The old score stays on the record, so you can see whether a model is improving. Each re-run has a cost cap so that no single model can distort our budget.
No. GenGrades does not publish an aggregate rating until there are real user votes behind it, and it never shows a number that was not measured. The only ratings on the site are ours, labelled as reviews by GenGrades.
Yes, that is the point. The prompts and input files for every test set are published, as are the raw outputs, watermarked. Run the same prompts on the same version and you should see the same kind of results; if you do not, tell us and we will re-test.