Breaking

GLM-5.2 Lands at #10 on Agent Arena, Top Open-Weight Result

June 17, 2026 at 16:22 EDT

Foundation Models
Open Source
Software Dev & Coding

Z.ai's (formerly Zhipu AI) GLM-5.2 (Max) has entered the Agent Arena leaderboard at #10 overall, marking the strongest open-weight result measured to date on the real-task evaluation platform. It posted a +9.4% confirmed task success rate and +14.9% on the Praise vs Complaint ratio against the baseline, rising three places from the previous generation GLM-5.1 at #13.

June 15, 2026 · Agent Arena Leaderboard

GLM-5.2 (Max) breaks into the agentic top 10 — the strongest open-weight result yet

Z.ai's MIT-licensed model lands at #10 overall on a leaderboard that scores real-world, long-horizon agent tasks — sitting just below closed flagships and narrowing the gap between open and proprietary systems.

#10

Overall rank — best open-weight entry to date

765,302

Sessions tallied across 27 models

~11K

Sessions evaluated for GLM-5.2 (Max)

Agentic coding climb: GLM-5.1 → GLM-5.2

Benchmark scores rebuilt as proportional columns — taller column = higher score.

Terminal-Bench 2.1

62.0

GLM-5.1

81.0

GLM-5.2

SWE-bench Pro

58.4

GLM-5.1

62.1

GLM-5.2

On FrontierSWE, GLM-5.2 comes within just 1% of the closed Claude Opus 4.8.

Arena metrics for GLM-5.2 (Max)

+9.43%

Confirmed Success (±4.52)

+14.88%

Praise vs Complaint (±9.11)

+4.37%

Net Improvement (±2.48)

1.86%

Tool Hallucination (±0.23)

What developers praise

Effective frontend leader in Code Arena
One-shot features for free that took time on paid models
MIT-licensed open weights · 1M-token context · 128K output

Trade-offs to watch

Steerability dropped -6.0% vs the previous model
Heavy resource demands for large-scale local execution
Further real-world evaluation still expected

Continue reading

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