Open weights vs Opus 4.8: one-shot wallet landing page test
Thread322 words
Same one-shot prompt to Opus 4.8, GLM-5.2, Kimi K2.7 and DeepSeek V4 Pro to design a wallet landing page. Opus and GLM were the best two, and GLM cost 5 cents against 27 for Opus.
Key claims
All four runs cost 40 cents in total
GLM-5.2 cost 5 cents, a fifth of Opus
DeepSeek V4 Pro ranked weakest
Can open weights out-design a closed frontier model?
@claudeai Opus 4.8 against @Zai_org's GLM-5.2, @Kimi_Moonshot K2.7 and @deepseek_ai V4 Pro.
Same prompt, one shot each: design a self-custody crypto wallet landing page in a single HTML file.
No edits, no retries.
Results in thread.
Same prompt to all four, run through @OpenRouter at temperature 0.7, one shot, no follow-ups.
Each model had to invent the brand, write the website copy, and build a custom hero animation, then return one self-contained HTML file.
All four together cost 40 cents to generate.
DeepSeek V4 Pro called its wallet Vault and opened with True wealth is self-sovereign over a floating gold gem.
Technically clean, but the most generic of the four, hero animation is sub-par, and it's a layout you have seen a hundred times.
It was also the cheapest run at 2 cents though.
Bottom of my ranking.
Kimi K2.7 named its wallet Nucle and led with "Own the chain. Not the anxiety." plus an orbital animation in the hero.
Competent, clean type, but nothing made me look twice.
The safe middle of the pack, and at 6 cents the priciest of the three non-Anthropic models.
Claude Opus 4.8 also went with Vault, but the execution was a level up.
"Your keys. Nobody else's hands." with a live particle sphere, real stats baked into the hero, and copy that read like an actual brand.
The catch is price: 27 cents, the most expensive run by far.
GLM-5.2 was the surprise.
It named the wallet Vellum and went editorial, almost print-like: Keys, kept like manuscripts. in a serif over a quiet radial animation.
The boldest art direction of the group, right next to Opus on quality.
Cost: 5 cents, a fifth of Opus for the same brief.
Verdict:
- Kimi was fine, DeepSeek the weakest.
- Opus and GLM were clearly the two best.
- GLM wins on cost/value.
Which one would you have shipped?
7 images on X