We tested Ox Alpha in Pi and DeepSeek Harness. Same task for both: use ego lite to scope out a few great sites, then build the most impressive single-file HTML landing page. The gap was bigger than we expected 👇

DeepSeek Harness. Hands down the standout, and honestly not what we expected going in. The final page had a level of polish we didn't see coming. It also took its time — this one clearly likes to think before it builds, and it ran way longer than Pi. Results:

Pi. Decent, but nothing that really stuck with us. The output was functional, just missing the finishing touches that made DeepSeek's version pop. It moved fast too — barely paused to think before diving straight into the build. Results:

Just a small tweak based on the prompt @kamellperry_ shared. Make the most visually stunning and coherent single HTML landing page about a random product of your choosing. First, use ego lite to find and study a few visually exceptional websites for inspiration. Make it in this worktree and really impress me. You can use anything you want to make it, it just has to be in one HTML file.

what are we doing here the first website displayed in the video is absolutely not AI generated :

This is the agent pulling up real websites for visual reference

Pi one is way better

Guess it's just personal preference.

try with omp

We dropped the prompt in the comments, feel free to try it out yourself.

Very interesting. DeepSeek harness will get a lot better

Ox alpha beating known models and still zero branding. Someone's enjoying the mystery. Was in this ai roundup

Same task, different harness, big gap. People should be measuring harness quality more than they currently do.

Same here, the difference in output quality was wild.

Pretty sure if you run it again there's a high chance results flip

Fair point

does the gap hold across repeated runs, because one landing page mostly measures sampling variance

ego lite 是否可以支持浏览器侧边栏标签？侧边栏的菜单 Tab,我似乎没有找到

We're currently testing vertical tabs and it'll be rolling out soon.

Is there any public benchmark available to compare DeepSeek harness with others?

DSH + Ox Alpha is Great

That single-file HTML page is a vanity project; the real bottleneck is the incentive structure driving the site selection.

Share the prompts na

Just a small tweak based on the prompt @kamellperry_ shared. Make the most visually stunning and coherent single HTML landing page about a random product of your choosing. First, use ego lite to find and study a few visually exceptional websites for inspiration. Make it in this worktree and really impress me. You can use anything you want to make it, it just has to be in one HTML file.

What about opencode? Can you test it too?

That single-file html constraint really forces the models to show their hand

Testing with ego lite is way too slow

Share prompt?

Make the most visually stunning and coherent single HTML landing page about a random product of your choosing. First, use ego lite to find and study a few visually exceptional websites for inspiration. Make it in this worktree and really impress me. You can use anything you want to make it, it just has to be in one HTML file.