General preference · Prompt adherence · Usability · Visual aesthetics
Contra Labs reports stronger results for GPT-6 Astra at the mockup stage, after a direction is established. Ideation, hero design, and text readability still need human judgment. Model preference and complete, working user flows deserve separate checks.
Read the source analysis| Rank | Model | Elo rating | 95% CI |
|---|---|---|---|
| 1 | Muse ImageMeta | 1626 | 1605–1647 |
| 2 | ChatGPT Images 2.5 FlareOpenAI | 1623 | 1597–1648 |
| 3 | ChatGPT Images 2.0OpenAI | 1551 | 1531–1571 |
| 4 | Nano Banana ProGoogle | 1546 | 1524–1568 |
| 5 | FLUX.2 [max]BlackForestLabs | 1457 | 1437–1477 |
| 6 | Grok Imagine 1.0xAI | 1391 | 1368–1414 |
| 7 | Krea 2 MediumKrea | 1306 | 1283–1329 |
Contra Labs Research · Redrawn from source data, not a live ranking or measure of artistic value.
Read complete research, from methods and figures to findings and limitations.
39 results
GPT 5.6 Sol led the first three evals and every loose brief. Once the work was staged across ideation, mockup, and refinement, the Claude models moved ahead.
Read articleSix models built landing pages for three products across ideation, mockup, and refinement. Opus and Fable finished first and second overall, but a different model led each stage of the work.
Read articleThe Human Creativity Benchmark moves to video: six models, three client campaigns, three phases each, judged head to head by working creatives. Seedance 2.0 won the set, and no model held the product together once it started moving.
Read articleNine designers annotated 20 landing pages one at a time, 738 notes in all. On Opus 5's pages the notes clustered in layout and readability, while brand fit and originality drew the fewest flags in the set.
Read articleWe scaled the Human Creativity Benchmark into a standing benchmark, starting with ad images: six models, three client campaigns, three phases each, judged head to head by professional creatives. Meta Muse Image won the set, and every model showed a signature failure.
Read articleAcross 600 blind comparisons and 400 write ups, designers liked the copy more than the rest of the page. Contrast, length and a few broken sections were what held it back.
Read articleFive expert product designers ran five Stitch iterations each on the same dashboard brief. The first prompts delivered wireframes and mockups in minutes; five rounds of refinement never raised the fidelity.
Read articleKimi K3 arrived ranked first on Arena's Frontend Code Arena. Across 480 blind comparisons by 8 designers, it finished level with GPT 5.6 Sol on landing pages, and edged ahead on the detailed briefs.
Read article8 designers annotated 40 pages from GPT 5.6 Sol, Claude Fable 5, Grok 4.5, and Muse Spark 1.1, marking 754 failure points across 8 tags. Every model broke differently.
Read article10 brand designers judged GPT Image 2, Nano Banana Pro, MAI Image 2.5, and Meta Muse across 12 logo briefs. The winner is clear, and still only half its logos cleared the client bar.
Read articleFive designers, four criteria, 6,400 blind pairwise ratings. Nano Banana 2 takes first on every brief-fidelity axis while the aesthetics standings invert behind it.
Read articleFive designers, five criteria, 8,000 blind pairwise ratings. Three of the four frontier models land within a few points of each other, and the best model changes depending on which dimension of visual quality you care about.
Read articleFour-model image battle across 10 briefs spanning portrait, environment, product, and lifestyle work, judged blind by working creative professionals.
Read articleGPT 5.6 Sol, Claude Fable 5, Grok 4.5, and Muse Spark 1.1 on the same ten landing page briefs, judged blind by 9 working designers as live interactive pages.
Read articleFour-model image battle across 10 briefs spanning the capabilities ByteDance advertises. Seedream won photorealistic generation at 35.7% and finished first or second in 57% of its tournaments.
Read articleFour-model style-transfer battle across 10 briefs. Muse took 29.5% of tournament wins and was the only model top-two on both wins and Elo.
Read articleClaude Fable 5 vs Claude Opus 4.8 on 5 real landing page and portfolio briefs, built in Claude Code, judged blind by 9 working designers.
Read articleA blind head-to-head of Seedance 2.0, Grok Imagine, Veo 3.1, and Adobe Firefly Video, judged by 12 professional video editors across 10 prompts.
Read articleTen professional designers ranked four frontier image models across nine dimensions of real design work. The models can make the work. Nothing on the market could reliably judge it, until we trained on the right data.
Read article10 designers, 4 models, 240 images. Spelling is solved. Typographic craft and client-readiness are where Ideogram v4 pulls away.
Read article11 production-style sessions. Gemini made the edit 73% of the time, kept the rest of the image still 64%, held both in 55%.
Read articleFour coding tools, 24 outputs, five working designers. The tool designers preferred to look at and the tool they'd put their name on turned out to be different.
Read article10 participants, 29 scored deliverables. The prompts that landed treated Gemini like a creative brief for a specific asset.
Read article8 targeted edits across 4 designer sessions. Firefly cleanly resolved 1, drifted on 5, and missed 2.
Read article4 designers, 3 deliverables each. The prompts that landed described physical direction, not aesthetic mood.
Read article5 designers, 5 openings, 1 luxury brief. The first prompt set what each session could reach.
Read article5 sessions, 5 designers, 1 real-world client brief. Strong as a starting structure, breaks under precision edits.
Read article42 sessions, 7 designers. ChatGPT Images 2.0 nails the typographic system, then breaks on the individual characters.
Read article41 sessions, 7 designers, 6 briefs. GPT Image 2 nails the concept, then breaks at production.
Read articleFour frontier image models, six brand campaigns, ranked blind by working creatives. GPT Image 2 wins the aggregate. Every other model owns a category.
Read articleFour-model style-transfer evaluation. Krea took #2 on style fidelity, 0.14 points behind GPT Image 2.
Read articleA blind head-to-head against the leading image models from Google, OpenAI, and Black Forest Labs, evaluated by professional creatives.
Read article12 models, 5 creative domains. One repeated complaint from working evaluators: the work all looks the same.
Read articleThe biggest phase-over-phase climb of any video model in the study. 3rd at ideation, 1st at refinement.
Read articleIdeation, mockup, refinement. AI fits differently at each phase, and the best creatives know where to hand off.
Read article61% win rate at ideation. 39% at refinement. The clearest model profile in our video evaluation.
Read articleHigher earning potential, more projects, no new hires. The survey from working independents.
Read articleof independent creatives report higher earning potential since adopting AI26% no · 8% other
Claude Opus 4.6, Gemini 3.1 Pro, ChatGPT 5.3 Codex, Qwen 3.5. The winner shifted at every phase.
Read articleSurvey of high-earning independent creatives. What they actually do with AI on real client work.
Read articleof AI output makes it to final deliverablesThe rest is stripped, reworked, or scrapped