Reviews
Same prompts. Every model. Scored blind.
A price tells you what a model costs, not whether it is any good. These reviews run one fixed prompt set through every model in a category, score the output without knowing which model made it, and print the bill.
No review is published yet. The bench is running its first wave, and a review appears here only after the run it reports on has finished. Nothing on this page is written ahead of the data.
What one review contains
Every time, in this orderThe same prompt set
Every model in a category gets the identical prompts. No per-model prompt tuning, because a tuned prompt measures the tuner, not the model.
A blind score
Outputs are scored under an opaque id, so the scorer cannot see whose answer it is reading. Where a task has a right answer, a machine check runs first and the scorer cannot overrule it.
The bill
What the run cost us, per model, next to the daily price from the index. A verdict without the cost of reaching it is an opinion.
An affiliate link never appears inside a ranking. One link per review, after the verdict, with the disclosure before it. Rankings sort mechanically, the same way the picks table does.
Categories and where they stand
Model counts from today's index| Category | Models | State | Why |
|---|---|---|---|
| Text generation | 605 | Running | Blind-scored answers, four tasks carry a machine-checked reference the scorer cannot overrule. |
| Image generation | 278 | Running | Five fixed prompts per model: portrait, hands, product, illustration, text-in-image. |
| 3D generation | 26 | Wave 2 | Needs an image input and a render step before a scorer can see the mesh. |
| Upscale and restore | 22 | Wave 3 | Needs a fixed input image set. |
| Image editing | 34 | Wave 3 | Same fixture work as upscale, cheap once it exists. |