Breaking

GLM-5.2 Ranks 2nd in Frontend Coding Eval, Tops Claude Opus

June 25, 2026 at 11:57 EDT

Foundation Models
Software Dev & Coding
Open Source

Z.ai's open-weight model "GLM-5.2 (Max)" has reached an Elo of around 1595 on "Code Arena: Frontend," the agentic frontend coding leaderboard operated by Arena.ai, placing 2nd overall. It sits behind the top-ranked "Claude Fable 5" (around 1654–1665) and surpasses several closed models including Anthropic's "Claude Opus 4.8 (Thinking)," sharply narrowing the gap with frontier models. Details

June 2026 · Z.ai (Zhipu AI)

Open-Weight GLM-5.2 Climbs to No. 2 on Frontend Code Arena

With a score of 1595, the MIT-licensed GLM-5.2 (Max) overtakes Anthropic's Claude Opus 4.8 on the community-voted Code Arena: Frontend board — trailing only Claude Fable 5, now just 70 points ahead.

1595

GLM-5.2 (Max) score — No. 2 overall

Points behind top-ranked Claude Fable 5

391k+

Blind votes behind the rankings

Code Arena: Frontend leaderboard

Bradley-Terry scores from blind, head-to-head votes on real web-dev tasks.

1665

Claude Fable 5

Closed · No. 1

1595

GLM-5.2 (Max)

Open · No. 2

< 1595

Claude Opus 4.8

Closed · now below

1408

GLM-4.6

Open · prior gen

One generation lifted the GLM series from 1408 → 1595 — a +187-point jump.

Where it shines

On par with / better than Claude for frontend & design
Clean React/JSX, strong visual quality
Stable over 20+ turns and long context
Substantially cheaper than Opus 4.8

Caveats noted

Not perfect; can over-block outputs
Best as a second opinion, not lead architect
Still trails top closed models on SWE-bench Pro
Latency in Max mode

Under the hood

744–753B

MoE params (~40B active)

1M tokens

Context · up to 128K output

81.0

Terminal-Bench 2.1 (Opus 4.8: 85.0)

MIT

Open-weight license

Continue reading

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