Two AI models rebuilt a Palace Museum painting site, and only one kept the details

What the test actually did

As Leiphone reported, its team set up a head-to-head test of two AI models: Google's new Gemini 4 Argon and the Chinese model GLM 5.3. One clarification on the framing: the models did not repaint a famous artwork. They rebuilt the Palace Museum's online painting-appreciation website, which holds over a thousand ancient paintings and offers zoom, filtering and detail pages.

The testers ran both models in OpenCode, a coding environment, with a reverse-engineering tool called REA and a companion skill called Reverse-skill. Each model first studied the live site, then wrote its own copy.

The scope was limited to the artwork display feature. Login, saving favorites, comments and the full annotation service were left out.

What REA is and why it is trending

REA stands for Reverse Engineer Anything. It is an open-source tool that gives an AI coding agent a way to inspect existing software. For desktop programs it can connect to analysis engines such as Ghidra, IDA and Hopper. For websites it can examine page structure, scripts, network traffic and screenshots, and it attaches evidence and known limits to its findings.

The tool was one of the hottest projects on GitHub in October. AI blogger Matthew Berman reacted with "Software is done." On October 16, Perplexity's CEO Aravind Srinivas said REA Skill was being added to its Computer product. Developers also showed off a near-complete rebuild of a Claude Slides editor using Codex and REA.

Leiphone also notes the pushback. Finding clues about how software works is not the same as rebuilding something that can be maintained for years. The article compares REA to an X-ray: it shows the skeleton, but the model still has to build the house.

How the two rebuilds compared

Gemini 4 Argon kept the original's skeleton. Its home page had the same three-column layout, four filters, a clear option, and titles, authors and sizes on each card. The header texture, fonts and spacing differed, and the paintings were swapped for geometric images, but the order of elements matched the original.

On the detail page, Argon's version kept the top banner, centered artwork information, a button to view details, and related recommendations. Buttons and text styling did not fully match, so the function and content were close while the finish was not.

GLM 5.3 built the same list, detail and viewing path, but changed the design more. Cards went from three columns to four, the light page turned brown, and the large filter area vanished. The four filters, top entry points and favorite icons on cards were all missing. The detail page swapped the original's big image and centered title for a white content box, though details and related items remained.

One odd touch: GLM 5.3 labeled its page as a minimal browsing version using local test images. Leiphone's verdict is that both are workable starting points, but Argon is closer to the original, while GLM 5.3 kept the browsing flow and rewrote or left out many details.

Where Argon stands on benchmarks

Leiphone pairs the test with a look at Argon's record. Google announced Argon on September 30 for trusted cybersecurity partners, then widened access on October 13 to its developer API and related subscription products. The focus is complex software engineering, professional knowledge work and multi-step agent tasks, meaning AI that calls tools across a long job.

On the Artificial Analysis intelligence index, Argon's high setting scored 53, tying GPT-6 Astra and Fable 5.1 at their max settings after rounding. Opus 5.5 scored 58 and Sonnet 5.5 scored 56.

Google's own results are mixed. On the AutomationBench knowledge-work test, Argon got 51.3% against 41.4% for Astra and 42.5% for Opus. On GraphWalks, a long-context test, its F1 score (a balance of accuracy and completeness) was 84.2%, against 71.8% for Astra and 66.8% for Opus.

Coding results swing more. Argon scored 77.9% on DeepSWE and 91.9% on Vibe Code Bench, ahead of rivals in the table. But it got 55.0% on FrontierSWE versus Astra's 65.5%, and 57.4% on Terminal-Bench 4.0 versus Opus's 66.4%. Leiphone concludes that Argon looks like a general flagship strong in knowledge work and long inputs, with competitive coding.

What it means for software makers

Leiphone argues the tools are leveling out while the models are pulling apart. REA lowers the bar for understanding a piece of software, and both models could take the site apart and rebuild it. The difference showed up in delivery: whether the copy respected the original's structure.

The outlet also suggests that simple display sites and generic features are getting easier to copy, which puts pressure on products that charge for features alone. Lasting value, it says, shifts toward data, services, compatibility and long-term upkeep.

Keep the limits in mind. This is one test by one outlet on one site, and the scoring was a visual comparison rather than a published method. A page that runs is not the same as software that can be maintained.