Introducing the Human Creativity Benchmark. The first eval that scores AI models the way creative experts do.

We ran today's leading models through real world >ideation >mockup >and refinement work and scored the output through the eyes of working creatives.

No model leads all three phases. @AnthropicAI's Claude leads ideation. @GoogleDeepMind's Gemini leads on design systems. @OpenAI's GPT closes hardest in refinement. Different stages of the work, different tools.

Two distinct signals show up in the data. >Convergence: where evaluators agree (prompt adherence, typography, CTA placement). This is teachable. >Divergence: where they disagree (mood, conceptual risk). This is taste, and it's still human

One concrete number: ad images that score 5 on usability finish top-2 in head-to-heads 84% of the time. Score-1 images do it 10% of the time. Usability is the strongest predictor of competitive success.

Link to paper:

Most people don’t yet understand the impact these systems are having on everything. Making the output clear and transparent like this, showing how/what and more importantly WHY? something ends up doing what it does is extremely important. I would love to see examples of evals on ad images ranked 1 vs 5, maybe some screen recordings on how these evals are playing out in head to head scenarios etc. Overall really love this. Want more.

thanks Linus🙏

Nice! Thanks for sharing these results. Would be so cool to see real-time screen record examples side by side. I need to try using Gemini 3.1 Pro Preview for some landing page stuff.

Pierre-Olivier Pelletier ( crushed it on this @contralabs_ai video. Featuring incredible work by lots of creatives on @contra including: • Sevd Atas • @therad_cactus • Cara Ellis • Elsa Toussaint • Blank Space Studio • Anita Autorino • @blakesteven • @desdivj • Mont Carver • @Rvoltedev

So proud of our team + our creative network who brought this research to life!

Congrats! Amazing work

my man

Build your own benchmarks is very smart congrats!

thank you ! more to come

Wow! Bullish on Contra Labs 🔥 I’m curious to know how many iterations it took before Contra labs was able to achieve this benchmark?

15,000 expert judgments and the conclusion is simple, AI can follow a brief perfectly and still produce work nobody would choose. Taste is the last human advantage. And it's not going anywhere.

This actually nails something most people miss convergence and divergence aren’t the same skill. AI is great at getting you to “good”, but getting to “distinct” still depends on taste. That’s where creatives still win.

So cool to be a part of this with you all!

This is dope NGL

Thanks Hasan

AI = speed Human = taste You need both to ship something great.

This actually matches what I’ve been feeling while working with AI. At the start, everything feels different lots of ideas, lots of directions. But once you refine it a bit, everything starts looking… equally “good”. That’s where it gets tricky. AI gets me to a solid version fast. But choosing what feels right? That’s still on me. So yeah, AI didn’t replace creativity for me it just made my taste more important.

The best way to use AI technically is by experiencing it first-hand. Creativity Benchmark is vital to ensure that this experience is 'right' Helping creatives and laymen to navigate their way to use AI models to their optimal efficiency is the smartest move. @contralabs_ai is doing this. It helped me restore my trust back on @grok and @imagine with their recent study - PROGRESSIVE ENHANCEMENT is a real thing. After a few prompts and phase of 'staging' the job, Grok helped me achieve better results.

damn this is literally a really important benchmark. good one!

This is amazing, really happy that humans are still considered xd. My creative process have been accelerated, more than anything. Brainstorming and researching through my niche with claude. Structure and softwares with chatgpt, and after that through iteration and trial in figma

How did you guys make this video? Is it all AI?

it is not, this was made by someone on Contra. All of the graphics are portfolio pieces from Contra users!

i love claude code for ideation and working on websites, gemini for images/nano banana, i love glif for experimenting, kimi2.6 on fireworks for pure speed, haiku for basic translation and transforms, it's just about finding the "pocket" of tools for the ai online utility belt.

love it when AI can speed up the process several times over🔥 I use Nano Banana pro, Claude and Runway - amazing tools

That’s sick

Interesting! I'm using a combo right now of ideation via chat, some agentic mockups and prototypes, plan mode to implement, then refine by testing and iterating. 📲

Why did you not include research tho For me Perplexity wins with research Gemini does well with fast ideation/wireframes and creating assets I use Chatgpt for refinements and prompt generation Not used Claude before, but believe it 'd be great based on what I ve heard

In my workflow, the best tool isn’t always the smartest one; it’s the one that lets me move from idea to execution without friction. AI has definitely sped up my process, but taste and clarity still feel like the real differentiators.

Most people are using AI wrong. No model dominates the full creative process. Some are for ideas, Some are for refinement If you’re using one tool for everything, you’re bottlenecking yourself. The advantage isn’t AI. It’s tool orchestration.

This is such valuable work, Ben. Love seeing Contra Labs actually ground the hype in real creative judgment instead of just benchmarks that miss the point.

Solid benchmark by @contraben. No model owns every stage Claude shines in ideation, Gemini in mockups, GPT/Grok strong in refinement.

"Creatives" are rabbits in Balenciaga fighting over the same bowl of carrots.

Lately I’ve been using AI a lot while working and learning. It really helps me go from an idea to actually building something much faster, and I feel more focused on the creative part instead of getting stuck.

Separating convergence (teachable: adherence/usability) from divergence (taste/steerebility) is super smart! Minor note: scaling the evaluator pool + adding real iterative loops would make it even stronger.

evaluating taste! congrats!

thanks Kate!

AI has helped me get to clarity faster as a UI/UX designer. @OpenAI: I use ChatGPaPT to explore ideas, refine UX copy, & shape case study narratives. @claudeai: I use it to break down briefs and structure user flows. @GeminiApp: I use it for research & competitor insights.

I run revenue for a DTC fashion brand through my Agentic System. Claude built the campaign strat, wrote emails, ran revenue analysis, and Klaviyo ops. Nano Banana 2 shot every product photo. I just steer. Open rates north of 60%, rev up 54% and the 'team' is two AI tools and me.

Very cool. Congrats on launching this Ben! Excited to explore

thanks Abdul