Design Arena is excited to introduce the company behind our work: Intelligence.
Over the past year, millions of people have helped us evaluate how well AI models perform on real-world creative tasks - where quality is subjective, outcomes are difficult to verify, and human judgment matters.
Design Arena is only the beginning. Let us know what you want to see next!
MAI Image 2.7 by @MicrosoftAI is now available on DesignArena!
MAI Image 2.7 is Microsoft's latest frontier model for generating and editing images, with finer control over color, material, and position at a fraction of the cost, averaging $35 per 1,000 images.
To test that control, we gave MAI Image 2.7 and MAI Image 2.6 the same crowded prompt: a vintage yellow tram with the number 541 painted on its side, stopping on a steep harbor street, with four specific passengers, a flower stall, and a round street cloc reading 7:15.
Only MAI Image 2.7 paints 541 on the tram's side. In MAI Image 2.7, the older woman holds the bunch of sunflowers, whereas in MAI Image 2.6 the sunflowers are placed in a seat. Neither image shows the clock at 7:15.
Congrats to the @MicrosoftAI team on the launch!
GPT-6.2 Astra by @OpenAI is now available on DesignArena!
GPT-6.2 Astra is OpenAI’s most capable model, built for long, hard projects, and OpenAI says it builds better websites, games, and 3D designs. It costs $8 per million input tokens, $40 per million output tokens, and $0.40 per million cached input tokens.
At launch, OpenAI showed GPT-6.2 Astra and GPT-6 Astra each planning a backyard vegetable garden. We stayed in the garden and gave both the same 3D prompt: a Victorian glasshouse on a gravel terrace at dusk.
GPT-6.2 Astra builds a grand glass hall with a tall dome and a long arched wing on each side. The hall features a far wider mix of plants, with palms fittingly placed in the center where the ceiling is highest. GPT-6 Astra builds a simpler glasshouse, and its lighting is more intense. GPT-6.2 Astra was also 3x faster and cost over a third less.
Congrats to the @OpenAI team on the launch!
BREAKING: Kling Image 4.0 by @kling_ai debuts at #13 on Image Editing arena with an Elo of 1247 and #14 with an Elo of 1279 on Image arena.
This puts Kling Image 4.0 in the same performance band as models from @OpenAI, @spacexai, and @GoogleDeepMind. At about $0.063 per image generation at 2K resolution, it matches Nano Banana 2.1’s quality for about 20% less per image. Kling also edits at 2K by default, and at about $0.066 per edit, you get 4× the pixels of most 1K editors for a similar price.
Congrats to the @kling_ai team on the launch!
Hy4 Flash by @TencentHunyuan is now available on DesignArena!
Hy4 Flash is Tencent’s new low-cost model for everyday productivity, at $0.070 per million input tokens, $0.281 per million output tokens, and $0.003 per million cached input tokens.
We gave Hy4 Flash and Hy4 the same 3D prompt: a mechanical keyboard that disassembles and reassembles by layer.
Hy4 includes more detail and realistic lighting, with its keyboard casting shadows that reflect its disassembly state.
Hy4 Flash is 8x cheaper and 2x faster. And surprisingly, Hy4 Flash built the mechanical keyboard people expect, with the standard QWERTY layout, whereas Hy4 opted instead for an ortholinear grid of same-size keys.
Congrats to the @TencentHunyuan team on the launch!
Static benchmarks only measure the time taken at runtime.
That single measurement doesn't reflect how long you wait when you actually use these models, where response time varies by time of day, and day of week.
On DesignArena, latency measured from real people running their own requests on their own time reveals how models hold up over a normal week.
Try making generations of your own and judge for yourself at DesignArena.ai.
Open-source models have a rush hour too, on Beijing time, but surprisingly they speed up briefly at lunch.
We looked at 40,000 requests to @Kimi_Moonshot's Kimi K2.7 Code, @Zai_org's GLM 5.2, and @MiniMax_AI's MiniMax M3, again sent through the lab's own API.
All three are fastest in the early morning and slow down sharply once the work day starts. Around 12:00 pm their speeds partly recover, and then they drop again to their slowest point of the day, around 3-4 pm.
Their swings are also much larger than OpenAl's. On weekdays, MiniMax M3 goes from about 19% above its weekly average speed before dawn to about 23% below it in the mid-afternoon, and Kimi K2.7 Code fluctuates even more, from roughly 47% faster at 5 am to about 25% slower after lunch.
As with OpenAl, weekends soften the dips. GLM 5.2 stays above its weekly average speed for nearly all of Saturday and Sunday.
Interestingly, this rush hour effect disappears on the weekends.
On Saturdays and Sundays, OpenAl's output speed stays about 6-11% above its weekly average all day. Even at its slowest weekend hour, around noon ET, it's faster than at any hour of the work week.
Anthropic's speed barely changes: weekdays and weekends both stay within about 2% of the weekly average, at every hour.
We control for response length, request type, and consider individual model versions.
Model latency significantly impacts user satisfaction. Surprisingly, @OpenAI's speed decreases noticeably during the work day, while @AnthropicAI's doesn't at all.
Almost all benchmarks measure latency at test time, but few look at how API response time performs live on production.
We looked at 200,000 LLM requests submitted by users on DesignArena, at every hour of the day, sent to each lab's standard API.
OpenAl model speeds decrease significantly during the work day, fluctuating up to 16%. Even more interestingly, Anthropic's models don't seem to be affected at all.
The @Alibaba_Qwen team has made huge improvements in Office work - this model is thorough:
In the promotion-interview template below, Qwen4-Max lists 2 failure stories while Kimi K3.1 pro tells a single professional failure
Congrats to the team!
New Qwen4-Max is officially the leading non-Western model in Slides and Spreadsheets on @DesignArena
It places 4th in Python-PPTX and 5th in Spreadsheets
In spreadsheets, this model is particularly strong at multi-turn calculation coherence, a surprising weakness in competing frontier models
Congratulations to the @Alibaba_Qwen team on this release
BREAKING: Qwen4-Max by @Alibaba_Qwen ranks #4 on Python-PPTX arena with an Elo rating of 1302 and #5 on Spreadsheet arena with an Elo rating of 1336.
On Spreadsheet arena, this marks a 30-position jump from Qwen’s previous model, Qwen 3.7 Max, and a 379-point Elo increase. On Python-PPTX arena, this marks a 5-position jump from its previous model and an 80-point Elo increase. This also places it in the same performance band as top models from @AnthropicAI, @OpenAI, and @Kimi_Moonshot.
Congrats to the @Alibaba_Qwen team on placing top-5 on both leaderboards!
Recraft V4.1 Flash V2 is live.
The fastest model on the market just got better: 1.1 seconds per image, now stronger on text in images and fine details. Same price.
Independently confirmed by @DesignArena: nothing on their board generates faster.
Try it: link in the replies ↓
With picture inputs, the video builds on what the user uploaded. The picture anchors the scene, which leaves the expanded prompt much less room to change it. Users picked Ultraquality 53% of the time, a coin flip.
Even the additional pixels didn’t seem to help, suggesting that prompt expansion, more so than post-training or resolution, might be what gives HeyGen its edge with text inputs.
Try making generations of your own and judge for yourself at DesignArena.ai!