Week 36, 2026 · August 31, 2026
Arena week 36: every model can spell now
All five models rendered FERMÉ LE LUNDI correctly, accent and all. Two years ago that would have been the whole story; this week it was the least interesting result on the board.
Text is solved at this length. Every model in this week's field — including the two-credit one — put the right letters on the chalkboard with the accent in the right place, which means the text probe has stopped separating models and now only separates eras. We are keeping it in the suite because the moment it starts failing again is worth catching, but the next revision will ask for a longer string.
What did separate them was the boring half of image generation: hands, fabric, and counting. Nano Banana 2 and GPT Image 2 took two matchups each and reached the top by opposite routes — Nano Banana 2 renders materials other models approximate, and GPT Image 2 does what it is told without embellishing. GPT Image 2 holds the higher rating on the order the results fell rather than on a better record, which is a twelve-point gap and not a verdict; a second week will say more about that than this one can.
The cost line is the one worth taking away. The text probe cost 11 credits on GPT Image 2 and 2 on Grok Imagine, and both got it right. Paying five times more for a chalkboard sign bought nothing this week.
Text rendering
Whether the letters the prompt asks for are the letters that appear, accents included.
The prompt
A hand-lettered chalkboard sign propped on the pavement outside a Paris café. The board reads FERMÉ LE LUNDI in white chalk capitals. Morning light, wet cobblestones, shallow depth of field.





The call
Every model got FERMÉ LE LUNDI right, accent included. Nano Banana 2 wins it for the sign nobody asked for: it put CAFÉ DU COIN · BRASSERIE on the awning behind and spelled that correctly too, which is the harder version of the same skill. GPT Image 2 drew the cleanest single line at five times the price.
Anatomy
Hands, mostly. Whether the fingers can be counted and whether they belong to the hands they are on.
The prompt
Close-up of a potter's hands centring wet grey clay on a spinning wheel. Both hands fully in frame, ten fingers, clay slip running over the knuckles. Studio window light from the left.





The call
GPT Image 2 is the only one that put both hands fully in frame with fingers you can count. Nano Banana 2 made the most photographic image by hiding one hand behind the clay, which passes the eye and dodges the question the prompt asked. FLUX.2 Pro merged the two hands where they meet.
Photorealism
Skin, fabric and light — and specifically whether a named material is rendered as that material rather than as an approximation of it.
The prompt
A fisherman in his fifties on a harbour wall at dawn, wearing a thick cream cable-knit wool sweater. Weathered skin, grey stubble, overcast light. Shot on 85mm at f/2.





The call
The prompt names a cable-knit sweater, and cable-knit is the part four of these five approximate. Only Nano Banana 2 renders the actual cable structure, and it threw in a legible boat name behind. GPT Image 2 has the best face in the set and the vaguest jumper.
Prompt fidelity
Counting and spatial relations. Whether 'exactly three' is three, and whether left and right are where they were asked to be.
The prompt
Exactly three green glass bottles standing on a white shelf. The tallest bottle on the left, the shortest in the middle, and a single yellow lemon on the shelf in front of the right-hand bottle. Plain white wall, soft even light.





The call
GPT Image 2 takes all four constraints — three bottles, tallest left, shortest middle, lemon in front of the right one — and adds nothing that was not asked for. Nano Banana 2 hits them too but invents branded labels. FLUX.2 Pro rendered four bottles for a prompt that says exactly three.
Where that leaves the board
| # | Model | Rating | Move | Won | Median cost |
|---|---|---|---|---|---|
| 1 | GPT Image 2OpenAI | 1591 | new | 2 of 4 | 11 credits |
| 2 | Nano Banana 2Google | 1579 | new | 2 of 4 | 6 credits |
| 3 | FLUX.2 ProBlack Forest Labs | 1443 | new | 0 of 4 | 2.5 credits |
| 4 | Grok Imagine ImagexAI | 1443 | new | 0 of 4 | 2 credits |
| 5 | Seedream 5 LiteByteDance | 1443 | new | 0 of 4 | 3 credits |
Cost is what the wallet actually charged that generation, not an estimate. There is deliberately no speed column: our generation records carry a start time and no finish time, so any duration we printed would be measuring our own polling loop rather than the model.
Run the winner yourself
Every model on this board is one dropdown in the same studio. Copy any prompt on this page, pick a different model, and see what you get.
Start picturing