Breaking

Z.ai's GLM-5.2 Tops Claude Opus 4.8 in Code Arena Frontend Ranking

June 26, 2026 at 13:58 EDT

Foundation Models
Software Dev & Coding
Open Source

GLM-5.2, an open-weight model from China's Z.ai (formerly Zhipu AI), has ranked higher than Anthropic's Claude Opus 4.8 (Thinking) in the frontend category of the community-voted Code Arena coding evaluation. GLM-5.2 (Max) sits at #2 with an Elo score of roughly 1,593–1,595, trailing Claude Fable 5 but surpassing Opus 4.8 (Thinking) at #3 (around 1,565) and other Opus variants. It ranks #2 in the React subcategory and #4 in HTML, and places among the top tier across several subcategories including Brand & Marketing, Reference-Based Design, and Data & Analytics. The latest scores can be checked on the leaderboard.

June 16, 2026 · Zhipu AI / Z.ai

An Open-Weight Model Cracks the Frontend Coding Top 2 — Above Claude's High-End Variants

GLM-5.2, Zhipu AI's MIT-licensed flagship, hit No. 2 on the community blind-vote Code Arena: Frontend board with an Elo near 1,594 — outranking Claude Opus 4.7 and 4.8 (Thinking), at a fraction of the price.

Code Arena: Frontend (blind voting)

744B

Total MoE params · ~40B active / token

Token context window

MIT

Open-weight license

Head-to-head win rates — GLM-5.2 vs. rivals

Share of blind-vote matchups GLM-5.2 wins. The 50% line marks an even split.

50% even

61.0%

Kimi-K2.6

59.4%

Claude
Sonnet 4.6

55.0%

Claude Opus 4.7
(Thinking)

45.5%

GLM-5.1
(tie)

Benchmark scores vs. Opus 4.8

Terminal-Bench 2.1

81.0 GLM-5.2

85.0 Opus 4.8

SWE-bench Pro

62.1 GLM-5.2 (still trails Opus on hardest tasks)

The price story

Frontier-class output, far below Opus

$1.40

input / 1M tokens

$4.40

output / 1M tokens

No. 1 among open-source models on the Artificial Analysis index — and routable into Claude Code for comparable results at lower cost.

Where it shines

Beats Claude in many frontend generation cases
Day-to-day usable as a frontier-class open model
Beat Claude Code on a Semgrep IDOR security benchmark
Runs standalone on a local Mac (256GB)

The caveats

Still trails Opus 4.8 on hardest long-horizon tasks
SWE-bench Pro remains a gap area
Some stayed cautious early as static benchmarks were limited

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.