跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

DeepSeek-V4-Pro Nails an Apple-Style Page Clone in One Shot but Fumbles a Pelican on a Bike

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agentic coding tools are being sold as general-purpose, but this test shows a sharp capability cliff: the same model that autonomously builds a production-grade marketing page can't draw a cartoon bird on a bike without human correction. Teams betting on AI-driven frontend work need to map tasks to model strengths, not assume uniform competence.

Summary

A hands-on test of DeepSeek-V4-Pro with the DeepSeek Harness agent reveals a model that is strong in uneven ways. Given a detailed, multi-thousand-word prompt to replicate Apple's AirPods Pro page — including browser research, responsive breakpoints, scroll-driven animations, and Playwright self-verification — the system produced a polished Next.js page in one shot with no manual fixes. The result was good enough to pass visual comparison against the real site.

That same model, asked to create a simple SVG animation of a pelican on a bicycle, generated a figure with missing handlebar grip and the wrong orientation. Even after self-correction via an image-recognition plugin, it missed obvious spatial errors until a human pointed them out. A side-by-side test with ByteDance's doubao-seed-evolving model showed the latter handled motion coordination and spatial relationships far more naturally on the same task.

The contrast points to a pattern: DeepSeek-V4-Pro excels at long-chain engineering tasks with clear constraints and repeated verification loops, but stumbles on open-ended visual problems that demand spatial common sense. The test also notes a steep price increase — from 4 yuan to 43 yuan per 100 million tokens — raising questions about cost-effectiveness for heavy usage.

Takeaways
DeepSeek-V4-Pro with Harness cloned Apple's AirPods Pro page — navigation, hero, sticky narrative sections, responsive layouts, and scroll animations — from a single prompt with no manual edits.
The system handled browser-based research, Playwright screenshot comparison, and two rounds of self-correction autonomously.
A simple SVG pelican-on-a-bicycle animation failed: the pelican's hands missed the handlebars and its body faced the wrong direction; the model needed human prompting to fix the orientation.
ByteDance's doubao-seed-evolving model produced a noticeably more natural pelican animation on the same prompt, with correct leg-pedal motion and spatial alignment.
DeepSeek-V4-Pro's token price jumped from 4 yuan to 43 yuan per 100 million tokens, roughly a 10x increase from the earlier Flash tier.
The Apple page clone achieved a 100% cache hit rate, suggesting the model may have leaned heavily on memorized patterns rather than pure reasoning.
Three test cases are not enough to generalize, but the pattern suggests the model is more reliable on constrained, verifiable engineering tasks than on open-ended visual generation.
Conclusions

The 100% cache hit on the Apple clone raises a question about how much of the output was reconstructed from training data versus reasoned from the reference page — a distinction that matters for copyright and originality in commercial work.

DeepSeek Harness's self-verification loop caught the missing handlebar grip but missed the orientation error, showing that vision-based self-critique still has blind spots a human spots instantly.

The price increase from 4 to 43 yuan per 100M tokens is steep enough to change the economics of agentic workflows that burn through tokens in long verification chains.

Doubao's stronger performance on the pelican task — a model the tester hadn't used in a long time — suggests the Chinese model landscape is shifting faster than casual observers track, with under-the-radar models catching up on specific capabilities.

Concepts & terms
DeepSeek Harness
An agentic framework that gives DeepSeek models browser access, tool use, and self-verification loops — it can open pages, take screenshots, run Playwright tests, and iterate on its own output.
doubao-seed-evolving
A model from ByteDance (Doubao) with 'evolving' in its name, suggesting iterative self-improvement; in this test it outperformed DeepSeek-V4-Pro on a spatial-reasoning animation task.
Cache hit (in LLM context)
When a model's output matches previously cached or memorized patterns rather than being generated from scratch; a 100% cache hit suggests the output may be heavily derived from training data.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗