DesignArena

New on DesignArena: MiniMax H3-Image by @MiniMax_AI is ranked #2 among open weight models on Image arena. This places it in the same performance band as top models from @ideogram_ai, @krea_ai, and @bfl_ai. At an average of 15s per generation, it is 4× than Ideogram 4.0 which is currently in 1st place and is the fastest model in the top 3. Congrats to the @MiniMax_AI team on a top-3 open weight placement!

Code is the hardest place on DesignArena to stay 1st. On our 15 Code leaderboards, the median run at 1st is 19 days. On Video it's 26 days, on Image 31, on Office 42, and on Audio 72.

Anthropic's biggest rival on DesignArena is Anthropic. Since July 2025, @claudeai models on our leaderboards have been dethroned 77 times. In 40 of those, the new leader was another Claude, usually a newer one. OpenAI's models took 1st from Claude 16 times, Moonshot's 10 times, and every other lab's 11 times combined. @OpenAI and @GoogleDeepMind are different. When one of their models was dethroned, it was usually by another lab's model: 20 of 28 times for OpenAI, and 21 of 29 for Google.

MiniMax H3-Image by @MiniMax_AI is now available on DesignArena! H3-Image is MiniMax's new open-weight model for making and editing images, with better prompt following and more precise editing. It's a 33.1B-parameter model that makes images up to 4K (4096 × 4096) without upscaling, takes up to 16 reference images, and costs about 2 cents an image on MiniMax's API. To test its editing, we gave H3-Image and Qwen Image Edit, another open-weight editor by @Alibaba_Qwen, an image of a workshop with paper models on a desk, and asked both to make it late evening with a brass desk lamp lighting the models. H3-Image turns the daylight to twilight blue and lights the desk with a lamp on a long brass arm. The shot stays as wide as the original, and the buildings outside the windows stay the same. Qwen Image Edit also adds a brass lamp, but it crops in closer and swaps the buildings outside for taller ones. Congrats to the @MiniMax_AI team on the launch!

Asked for a Charminar for Blender, Fable 5.2 built one you can actually export, as GLB or OBJ, with Blender import steps. It also nails the details like the four working clocks, the rooftop mosque, the lotus-based minaret domes, as well as birds flying around in the sky. Inkling 2 Small set a nice scene at a road crossing, but added a central dome the real building doesn't have, and gave no way to get the model into Blender.

In the example below Fable 5.2 turned the prompt into a fully thought-out scene: candlelit paths, customizable snowy weather, a manger, and a bell tower that actually rings, with greetings in the languages of the countries that celebrate it. Kimi K3.1 Pro built a tidy daytime diorama, but its candles start unlit and it has no bells at all. Thought it has snow, it is difficult to see against the white background it is set in.

New in DesignArena: Claude Fable 5.2 by @AnthropicAI lands at #1 on 3D Design Arena with an Elo of 1540 and 81% win rate, replacing GPT-6 Astra (max) by @OpenAI. The second best Anthropic model on 3D Design arena is Opus 5.5 in 3rd place, which is followed by Inkling 2 Small by @thinkymachines and Kimi K3.1 Pro by @kimi_moonshot in 4th and 5th place respectively. At $8/$40 per 1M input/output tokens, it falls within a similar cost per generation as Claude Opus 4.6 while having a much higher Elo. It excels at creating detailed 3D design settings that are well-suited to each user's request. Take a look at some examples below:

Claude Fable 5.2 by @AnthropicAI is now available on DesignArena! Claude Fable 5.2 is Anthropic's new model for hard coding, reasoning, and vision work, at 20% less per token than Fable 5.1 ($8/$40 per million tokens, down from $10/$50). It can read up to 1M tokens, write up to 128K, and has a new fast mode that writes up to 2.5× faster. We gave Fable 5.2 and Fable 5.1 the same two 3D prompts, a brass orrery and a chameleon at twilight, to see what changed. Fable 5.2's orrery has engraved month and zodiac names, warm reflections that shift across the brass as the camera moves, and a photoreal Jupiter. Fable 5.1's orrery is built from simpler, chunkier parts under even lighting with few reflections, and its Jupiter is a plain striped ball. Fable 5.2's chameleon has lifelike colors and patterns, moss that glows brightly under the branch, and fireflies that pulse. Fable 5.1's chameleon is made of rough, blocky shapes, its moss gives off only a faint glow, and its fireflies' pulsing barely noticeable. Congrats to the @AnthropicAI team on the launch!

In the example below, Muse Ember generated an aloe plant with lifelike leaves as requested, whereas Muse Wonder Max generated thin, flat leaves. Muse Ember also added white spines that line the leaves of real aloe and was over 50× cheaper and 4× faster.

In the example below, Muse Ember was 25× faster at generating 5 different tree models at 33 seconds, with each tree getting its own label. Muse Ember also provided more granular details per tree including height, number of polygons, and a wireframe view. Comparatively, Kimi K3.1 Pro took 14 minutes and generated 5 similar looking models.

BREAKING: Muse Ember by @AIatMeta has redefined the Preference vs Speed Pareto frontier on Overall HTML, knocking out 10 top models from @GoogleDeepMind, @OpenAI, and @DeepSeek_ai. At 33s per generation, Muse Ember is 9.5× faster on average than the previous Pareto frontier models. It is also at least 34× cheaper on average than all the previous Pareto frontier models it replaced at $0.10/$0.50 per 1M input/output tokens ($0.004 cache). It is also ranked 10th overall with an Elo of 1345! A huge congrats to the @AIatMeta team on a top-10 placement on the leaderboard!

New on DesignArena: Gemini Omni 1.2 Flash by @GoogleDeepMind is projected to rank #3 on Video arena. That would give Google the top three spots on the leaderboard, dominating the Video generation field! It trails only Gemini Omni 1.1 Flash and Gemini Omni Flash, and is projected to finish ahead of Wan 3.0 by @Alibaba_Wan and P-Video-2-Pro by @PrunaAI. Official results pending. Congrats to the @GoogleDeepMind team!

Ultimately, no AI assistant is as good as a human personal assistant yet. LifeArena is run on Gmail inboxes with 30K unread messages, years of text messages, and hundreds of personal contacts. More details on methodology coming soon. In the meantime, let us know if there are more agents you'd like us to try!

People brought their messy problems: breaking a lease shared with a roommate in Seattle, disputing a tow in San Francisco, and orchestrating a $2,500 Christmas trip to Florida (we're jealous). #1 @Muse by @Meta was the strongest all-around; it was fast, great at displaying images when necessary, and never pestered users with unnecessary follow-up questions. #2 Dots by @OpenAI was quite strong at computer use and completing tedious admin tasks. It fell short on coming up with thoughtful gifts, enticing vacations, and anything that required depth in user taste. #3 @instinct was a very convenient form factor. But it was, at times, unbearably slow, an issue that seems to have ballooned with the platform's popularity. #4 Grok Bot by @SpaceXSI struggled on start. Despite having access to gigabytes of user PII, it failed to build the same depth of personal understanding unprompted as its competitors. #5 Gemini by @Google was the least autonomous. It requested permissions to the point of becoming infuriating for users, declining to complete purchases even when instructed to do so.

Introducing LifeArena: the world’s first arena for personal assistants, as judged by real users. As AI assistants become more capable, existing static benchmarks struggle to capture the creative, unthinkable requests of people in their everyday lives. So, we offered select users the most popular personal assistants: @instinct, @Muse by @AIatMeta, Dots by @OpenAI, Grok Bot by @xai, and Gemini by @GoogleDeepMind. The results are in: @Muse is #1... by a landslide. Here's what users said:

In the example below, the user asks for a gymkhana rally car drifting a continuous donut in the street. Inkling 2 Small frames the drift with a cinematic chase cam that stays low behind the car as the smoke rolls toward the lens. The street around it is a detailed rally environment, with neon storefronts, string flags, crowds along the course, and traffic cones placed around the tire stack the car circles. LED lights under the car match the neon glow, and four camera angles include a curbside view at road level. Kimi K3 holds a clean donut, layers its tire marks, and gives its car racing stripes and a number. However, it places the drift in an empty gray plaza, with the cones bunched in the middle of the circle, and its chase cam watches from above.