September 8, 2026
September model refresh
We're now carrying 120 models across image, video, audio and 3D. This month we refreshed the defaults and re-checked every price.
A new frontier model shows up roughly every fortnight. A catalog nobody curates is just a list. So every month or so we go through the whole thing. We add what is new, retire what has been beaten, and re-check what each one costs us.
New and worth trying
| Model | What it's good at |
|---|---|
| Gemini 3 Pro Image | Following a complex prompt, and text inside an image |
| Gemini 3.1 Flash Image | Fast iteration while you're still exploring |
| Seedream v5 Pro | Photographic product and fashion work |
| FLUX 2 Flash | Quick concepting and layout |
| Qwen Image 3 | Illustration and stylised work |
| Grok Imagine 2.0 | Expressive, character-forward images |
Auto got smarter, not just bigger
More models only helps if the default choice is a good one. Auto now refuses to hand your prompt to a model that ignores reference images when you've actually attached references. It moves to one that can use them, and it tells you it did. The old behaviour was to drop them silently, and leave you wondering why the result looked nothing like your product.
Every model shows what it costs before it runs. Three numbers matter here: the estimate in the composer, the figure the agent asks you to approve, and the charge on the finished result. All three read from the same table, so they cannot disagree.