Arena.ai gave the same set of frontend coding prompts to GLM-5.2 (Max) and Claude Opus 4.8 (Thinking) for single-shot generation and recorded the results side by side. The comparison shows that GLM-5.2 (Max) ranks higher than Opus 4.8 (Thinking) in Code Arena: Frontend community voting, and that these prompts can be copied directly for testing frontend visualizations or React site generation.
Suitable tasks: HTML data visualization (magazine-style charts, radar screens, timelines and stories); single-page visual stories; React frontend sites (themes such as a ramen shop and a record pressing plant); frontend tasks built around “single-shot generation, self-contained output, and source footnotes.”
Unsuitable tasks: workflows that require multiple rounds of iteration and user feedback (this set is a single-shot comparison); frontend tasks that require a backend or database (the prompts explicitly require a frontend-only implementation with local seeded data).
Applicable model versions: GLM-5.2 (Max tier); comparison model Claude Opus 4.8 (Thinking tier).
Applicable clients, Agents, or APIs: the Arena.ai Code Arena: Frontend task environment; any coding Agent or API that supports single-file HTML/React output.
Recommended reasoning tier and parameters: use GLM-5.2 in Max tier (the official coding recommendation); no special sampling requirements, complete in a single shot.
The following seven prompts are the complete original texts retrieved after being expanded in the thread (the thread claims there are 10 examples in total; the remaining examples were not fully expanded within the visible range). Every task requires “using only the supplied frozen CSV/local seeded content, keeping the result self-contained, and adding a small source footnote.”
Create a beautiful HTML day-in-the-life visualization from this hourly activity data. The result should feel closer to a magazine graphic than a dashboard and should make the rhythm of the day intuitively readable. Use only the supplied frozen CSV. Keep the result self-contained and add a small source footer.Build an HTML visual explainer comparing the seasonal Arctic sea ice cycle across selected years. Make the annual loop or seasonal shape the central visual idea and keep the climate signal easy to grasp at a glance. Use only the supplied frozen CSV. Keep the result self-contained and add a small source footer.Create an HTML chart or one-slide visual story about the number of objects launched into outer space each year. Find a visual treatment that makes the recent break from the historical pattern feel dramatic but credible. Use only the supplied frozen CSV. Keep the result self-contained and add a small source footer.Create an imaginative HTML visualization of near-Earth asteroid close approaches during 2025. It can borrow the visual language of a radar screen or orbital diagram, but the distances and velocities must remain understandable. Use only the supplied frozen CSV. Keep the result self-contained and add a small source footer.Build a compact HTML visual explainer showing how Nobel Prize categories and laureate counts evolved by decade. Aim for a clear editorial overview with enough texture to reward closer inspection. Use only the supplied frozen CSV. Keep the result self-contained and add a small source footer.Build a bright, polished React website for a beloved neighborhood ramen shop on a busy Friday night. Use seeded local content only. The site should open with a full service already underway: orders on the rail, broth timers running, seats turning over, toppings being prepared, and a few pickup tickets nearly ready. Visitors should be able to explore the menu, build a bowl, choose a seating or pickup window, watch how their order moves through the kitchen, and see the wait time and price change as choices are made. The design should feel warm, energetic, detailed, and delicious without relying on restaurant photography.Create a front-end-only React website for a boutique vinyl record pressing plant handling limited-run album releases. Seed the site with an album already in production: lacquer approval, stampers, vinyl color, sleeve assembly, test pressings, release bundles, and shipping windows should all be present on load. Visitors should be able to follow the pressing run, switch vinyl variants, inspect packaging, reserve a numbered copy, and see the edition count and ship date update. The design should feel bright, analog, tactile, musical, and precise.Same prompt, single-shot completion; the two models’ outputs were compared in screen-recorded videos (each reply in the thread contains two videos).
GLM-5.2 (Max) ranks No. 2 overall in Code Arena: Frontend, scoring +29 points higher than Claude Opus 4.7 (Thinking) and trailing only Claude Fable 5; it ranks No. 2 on the React sub-leaderboard and No. 4 on the HTML sub-leaderboard, making it the highest-ranked open-source model.
Community views in the thread’s comments (for reference): a single-shot head-to-head can show that a lower-ranked model may still win on specific tasks; some also questioned whether it was simply fortunate on these particular prompts.
The prompts’ “frozen CSV” or “local seeded content” is part of the task input, but the thread does not make these data files public; you need to supply your own data when reusing the prompts directly.
The comparison reflects the Max/Thinking tier configurations as of 2026-06-27 and does not apply to other GLM-5.2 tiers or later versions.
GLM-5.2