Mistral just shipped Large 4, its biggest model ever. So I gave the same 3D prompt to 8 models. One shot, no fixes. Open: Mistral Large 4, Kimi K3, GLM 5.3, DeepSeek V4 Pro Closed: Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol Results below 👇

My take: Eiffel Tower: Mistral's is the best of the open models. GPT-6 Astra wins overall. Aquarium: Mistral is near the bottom. Opus wins. The Chonk: Mistral's cat is fine, GLM and Kimi do better. All 3 scenes cost Mistral $0.87. Opus: $13.83.

One thing to know if you try it: Mistral Large 4 has reasoning OFF by default. My first run with it off: the Eiffel Tower came out as a black screen. Turn it on (reasoning_effort: high) and it's a different model. Slower though: 66 min for the 3 scenes.

D'après leurs annonces, ce modèle est optimisé pour l'administratif, la compliance, la cohérence documentaire,etc. C'est un modèle destiné au BtoB. Je ne sais pas s'il existe des benchmarks sur ces domaines précis mais on verra bien quand il sera déployé en entreprise.

yes je suis en train de tester ça sur des projets où j'avais GLM pour comparer

This kind of real-world testing actually shows what matters most. We looked at Large 4's positioning in the capability race here.

Looks to be very slow based on your results

Yes, it was quick without thinking, butthe results were terrible

Open weights are the real flex when they're this close to the closed ones.

one shot with no fixes is a fair test, but the gap i care about shows up in round 3, when you ask for a change and see which model breaks what already worked

This is Funny, every "good" bench came from French user...👀

bet claude still took it