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

It's been a few days since DeepSeek-V4-pro and DeepSeek Harness were released, and I haven't tested their actual performance yet.

DeepSeek has officially announced a price increase, so I'll test it right when the price goes up. Don't be afraid, don't back down — we've got money, we're just that bold.

Let's test it out today.

But let me be clear upfront: this isn't a benchmark. It's an experiential test using a few real-world tasks. Different cases have different complexity, context length, and tool calls, so the judgments below are only valid for these specific results and shouldn't be taken as a model leaderboard.

I'm using DeepSeek-V4-pro at max effort intensity, in standard mode.

The first example is to test the pelican effect.

Create an HTML file with a 2D SVG animation of a pelican riding a bicycle.

Hmm, this pelican looks a bit abstract. Its hands aren't holding the handlebars, and the pelican's orientation is wrong.

I installed an image recognition plugin for Harness and had it analyze the problem itself.

It did notice that the hands weren't holding the handlebars, but it overlooked what I thought was an obvious issue — the pelican's direction was wrong.

After I pointed out the direction issue, it discovered and corrected it. I made a GIF to show everyone.

Why does this pelican like old-school bikes so much... And what are those little hands underneath? Paper-mâché claws? Why do they bend, and why are they sticking out from below...

After this test, my stomach suddenly started hurting, so I went to the bathroom.

(15 min later...)

When I came back, I suddenly saw a message from Doubao. I thought, Doubao? That familiar yet unfamiliar name.

I haven't used Doubao's models in a long time. I remember when I first used Ark's coding plan, I was using Doubao's model. So what does it want from me today? I never activated these services...

So I logged onto the official site and found that Doubao has a new model called doubao-seed-evolving, which I hadn't heard much about before.

Evolving? Sounds like evolution? Has Doubao evolved?

Out of curiosity, I looked into the doubao-seed-evolving model.

Is it a mule or a horse? Let's take it for a spin. So I did a pelican test with this model.

I didn't have high expectations, but this pelican surprised me a bit.

Aside from being a bit skinny, this pelican has no real issues. What I really like is how its legs follow the pedals in circular motion — this is the most coordinated test result I've seen. Let me show you the GIF version.

(This pelican working so hard reminds me of student days, munching snacks while frantically pedaling home every... evening after school.)

Honestly, the pelican from doubao-seed-evolving turned out really well. I didn't expect it to be this smooth — at first glance, I couldn't find any obvious flaws.

At least it's better than the V4 version above. That pelican looked like a retired old man riding an antique bike with a radio playing the news beside him.

If I had to nitpick, maybe the bike chain isn't moving, like coasting in neutral at a red light.

Or maybe the pelican isn't refined enough?

At least in this test, doubao-seed-evolving handled motion coordination and spatial relationships much more naturally.

Suddenly this model seems interesting. I'll test it more later. Today's main quest is DeepSeek-V4-pro.

The next test: have DeepSeek-V4-Pro make a game.

Specifically, the Prompt War game I made before.

First, let's look at the effect I achieved previously.

Now the DeepSeek-V4-pro test result with the same prompt.

This time, DeepSeek's aesthetics aren't as good as my previous version, but the smoothness, completeness, and experience are decent.

One issue: the cache tower effect is supposed to slow things down, but when I actually played it, there was no slowdown.

For the next test, I had DeepSeek-V4-Pro replicate an Apple website. I wanted to see DeepSeek-V4-Pro and DeepSeek Harness's capabilities in browser observation, visual deconstruction, frontend implementation, scroll animations, responsiveness, and self-verification.

Here's the prompt:

You are a senior frontend engineer, interaction designer, and visual restoration expert.

Please visit and study the following reference page:

https://www.apple.com/airpods-pro/

The goal is to implement a high-fidelity Apple-style AirPods Pro product showcase page from scratch locally, for internal model capability evaluation.

Don't just make an "Apple-like" generic landing page. You need to accurately restore the layout, visual hierarchy, responsive behavior, and scroll experience based on the currently visible reference page.

I. Replication Scope

Only implement the following sections — no need to replicate the entire apple.com:

1. Apple-style global navigation bar
2. AirPods Pro product navigation bar
3. Hero section above the fold
4. At least 3 consecutive product feature sections
5. One narrative section with sticky or scroll-driven effects
6. Product selling point or spec card area
7. Purchase CTA area
8. Apple-style Footer

Final page route:

/airpods-pro

II. Reference Page Research

Before coding, use the browser to complete the following:

1. Open the reference page at three sizes: 1440×900, 1024×768, 390×844;
2. Screenshot the hero and main sections;
3. Record the page's:
   - Maximum content width
   - Font sizes and weight hierarchy
   - Main background and text colors
   - Section spacing
   - Image aspect ratios
   - Border radius
   - Sticky areas
   - Scroll animation trigger methods
   - Mobile layout changes
4. Before starting implementation, summarize the page structure and design patterns in a short checklist.

Directly copying the original site's HTML, CSS, JavaScript, React components, or build artifacts is prohibited. You may only reimplement based on what is visible in the browser.

III. Technical Requirements

If the current directory already has a frontend project, continue with the existing tech stack.

If the current directory is empty, use:

- Next.js
- TypeScript
- CSS Modules or Tailwind CSS
- Prefer native CSS for animations
- Only introduce animation dependencies when truly necessary

Don't introduce unnecessary production dependencies just to complete the task.

The page must:

- Be locally runnable
- Have no TypeScript errors
- Have no obvious console errors
- Support desktop, tablet, and mobile
- Have smooth animations
- Support `prefers-reduced-motion`
- Not depend on the original site's private APIs
- Not hotlink Apple's image resources

Product images can use self-made simplified AirPods diagrams, CSS graphics, or local placeholder assets. For internal evaluation, publicly displayed reference images may be downloaded locally, but must be centralized in `public/reference-assets/`, with sources recorded, and the page must not depend on remote resources. If the page is to be published publicly, Apple's branding, logo, copy, and product assets must be replaced.

IV. Visual Requirements

Focus on restoring the following characteristics:

- Large black or high-contrast backgrounds
- Oversized product titles
- Precise font weights and line heights
- Restrained color usage
- Generous whitespace
- Scale relationship between product images and text
- Apple-style rounded content cards
- Sticky product showcases
- Titles, descriptions, and product visuals that progressively appear during scroll
- Large-canvas feel on desktop
- Mobile layouts that are re-composed, not just scaled-down desktop versions

Avoid:

- Common SaaS template layouts
- Blue-purple gradient backgrounds
- Excessive shadows
- Arbitrary icon additions
- All elements fading in simultaneously
- Using the same card structure for every section
- Using emoji instead of formal graphics
- Adding visual elements not present on the reference page just to "look premium"

V. Interactions and Animations

At minimum, implement:

1. Navigation bar scroll state changes
2. Hero content entrance animation
3. One sticky product showcase section
4. Subtle product image scale or displacement animations
5. Text appearing in stages with scroll
6. CTA button hover and focus states
7. Mobile navigation interaction
8. Disabling complex animations in Reduced Motion mode

Animations must be driven by page content. Don't add flashy effects unrelated to the reference page.

VI. Verification Process

After implementation, you must:

1. Start the project;
2. Run lint, type checking, tests, and production build;
3. Use Playwright to visit:
   - 1440×900
   - 1024×768
   - 390×844
4. Screenshot the full page and key sections;
5. Visually compare implementation screenshots with reference page screenshots;
6. Perform at least two rounds of proactive corrections;
7. Check:
   - Font hierarchy
   - Page width
   - Section height
   - Image dimensions
   - Alignment relationships
   - Mobile overflow
   - Sticky animations
   - Page jank
   - Horizontal scrollbars
   - Console errors

Don't stop after the first successful run. Visual comparison and proactive correction are part of the task.

VII. Acceptance Criteria

The final result should meet:

- Recognizable as an Apple product page style at first glance;
- Hero, product images, titles, and whitespace proportions close to the reference page;
- Desktop and mobile versions independently designed;
- At least one complete scroll-driven narrative section;
- The page is not a static screenshot — content and interactions are real;
- No obvious layout overflow;
- Animations smooth and not excessive;
- Project can be started with a single clear command;
- Tests and production build pass.

VIII. Final Report

After task completion, provide:

1. Project start command;
2. List of implemented pages and features;
3. Technical approach used;
4. Test and build results;
5. Playwright screenshot paths;
6. Remaining differences from the reference page;
7. Specific adjustments made to correct visual differences;
8. List of main files modified.

Please autonomously complete the research, implementation, startup, testing, screenshot, and correction process.

Folks, after a long wait, the big one has arrived.

This was delivered in one shot — I made no modifications.

If you don't believe it, you can test it directly with the prompt above.

(How the hell did they achieve 100% cache hit...)

What surprised me most about this result is precisely how it contrasts with the pelican earlier.

The pelican seemed simple but heavily depended on spatial relationships, motion coordination, and visual common sense.

Replicating the Apple page is far more complex, but the reference target is clear, the requirements are finely decomposed, and Harness can continuously observe and verify through the browser.

At least from these few cases, DeepSeek-V4-Pro performs more stably in long-chain engineering tasks with clear constraints and repeated verification than in open-ended visual problems.

Of course, three cases aren't enough to prove this is a universal pattern for the model, but it does explain why the earlier test felt a bit abstract while this task suddenly changed my impression.

Really, this AirPods replication test genuinely changed my view of DeepSeek Harness + DeepSeek-V4-Pro.

I vaguely remember cursing DeepSeek Harness the hardest when it was released...

But I still have to curse about DeepSeek's price increase.

I haven't tested other models on Apple website replication yet, so I can't conclude whether DeepSeek offers good value for money.

But it's genuinely expensive — it nearly bankrupted me.

I still remember when DeepSeek-V4-Flash cost only 4 yuan per 100 million tokens. Now 120 million tokens cost 43 yuan...

After this round of testing, my judgment on DeepSeek Harness + DeepSeek-V4-Pro is:

It has achieved a surprising one-shot delivery capability on complex frontend tasks requiring browser observation, long-chain execution, and self-verification; but on simple SVG animation tasks that depend on spatial relationships and motion coordination, stability is still insufficient.

Simply put, it's not uniformly strong or weak — it's strong, but unevenly so.