Try beating both yourself at designarena.ai/judge/flying-…. Can you tell which is which?

In the other, the car keeps tumbling, so you have to time the burn for when the wheels swing back under you. Its rings also sit on the car's actual flight path. With no input at all, it flies through all three.

However, we were surprised to find that in one game, the car only spins while you're steering it. With no input, it stops rotating in midair, making it easy to level out and land. Its bonus rings can't be reached either. They sit about 90m up, and the car tops out around 67m.

We read both games' code and replayed them with identical inputs, frame by frame. Both get the basics right. They count flips, run on a limited tank of fuel, and aim the thrust wherever the car is tilted, so it only slows your fall when the car is level.

Two models built the same game inspired by the @Tesla Roadster. One of them "broke" physics to make it easier. Can you tell which is which? A user asked DesignArena for a game about landing a flying car. @Alibaba_Qwen's Qwen 3.8 and @OpenAI's GPT-6.1 Sol each built one.

Fibo 2.0 by @bria_ai_ is now available on Design Arena! Bria’s new 8B-parameter open weights image model delivers strong performance across text-to-image generation and image editing. It brings major improvements in text rendering, layout, and precise text edits, with API pricing starting at $32 per 1,000 images. Congrats to the @bria_ai_ team on the launch!

Try beating both yourself at designarena.ai/judge/flying-…. Can you tell which is which?

In the other, the car keeps tumbling like a real object, so you have to time the burn for when the wheels swing back under you. It's harder, but it's more real. Its rings also sit on the car's actual flight path. With no input at all, it flies through all three.

However, we were surprised to find that in one game, the car only spins while you're steering it. With no input, it stops rotating in midair. That makes it easy to level out and land, but real cars don't do that. Its bonus rings can't be reached either. They sit about 90m up, and the car tops out around 67m.

We read both games' code and replayed them with identical inputs, frame by frame. Both get the basics right. They count flips, run on a limited tank of fuel, and aim the thrust wherever the car is tilted, so it only slows your fall when the car is level.

Voxtral TTS 3 by @MistralAI is now available on Design Arena! Voxtral TTS 3 powers real-time voice agents with support for 35+ languages, natural turn-taking, and sub-second response times. Its streaming pipeline combines speech recognition, reasoning, tool calling, and speech generation, enabling agents to handle tasks like identity verification, billing inquiries, and appointment booking across phone calls, browsers, and apps. Congrats to the @MistralAI team on the launch!

3/ Even without reference images, Krea 3 builds complete scenes from a prompt alone: models styled head to toe in full outfits, set in moody, cinematic lighting.

2/ Krea 3 can take a fabric swatch and apply it to any object, and it understands how that object is built. Seams, cushions and folds all follow the shape, which is essential for realistic product mockups like the one below.

1/ Give Krea 3 a reference image and it carries it onto a model with full fidelity, holding every detail through dynamic poses, new camera angles, and shifting light.

BREAKING: Krea 3 (Ultra Thinking) by @krea_ai lands at #4 of 51 on Design Arena's Image Editing leaderboard, making Krea the #2 Image Editing lab after @OpenAI! Krea 3 (Ultra Thinking) edges out Reve 2.0 by @reve and climbs even higher in key subcategories, ranking #2 in Architecture and #3 in Landscape. Congrats to the @krea_ai team on the launch!

Great question! The current length is still 5 seconds, but we have updates coming soon :)

New on Design Arena: DeepSeek-V4.1-Flash-1017 by @deepseek_ai is projected to rank #3 among open-weight models on our Website leaderboard. This marks a 16-position jump past DeepSeek-V4-Flash-0731 among open-weight models. It trails only Kimi K3 by @Kimi_Moonshot and MiMo-V2.6-Pro by @XiaomiMiMo, and is projected to finish ahead of GLM-5.3 by @Zai_org and Ling-3.1-flash by @AntLingAGI. Congrats to the @deepseek_ai team!

Watch the full animated video here: piped.video/tGE8D6xwCSU

You may think the frontier models structure their thinking somewhat similarly. Our research found that their generation process differ wildly, and you can see it in the output. Watch as we compare @OpenAI's GPT-6 Astra and @Anthropic's Claude Opus 5.5 nitter.siddz.com/i/article/210613619273…