Z.ai's latest model, GLM-5.2 (Max), has entered the new Agent Arena benchmark at No. 10 overall and ranks first among open-weight models by a wide margin. Its score sits close to Claude-Opus-4.8 (non-thinking), as shown on the Agent Arena leaderboard on June 16, 2026.
June 16, 2026 · Z.ai (Zhipu AI)
GLM-5.2 cracks the Agent Arena Top 10 — and leads every open model
Z.ai's open-weight contender now ranks No.10 overall on the real-world Agent Arena leaderboard, matching Claude Opus 4.8 (non-thinking) on agentic tasks while costing a fraction as much — and it launched with no traditional benchmarks at all.
No.10
Overall rank on Agent Arena — up from ~No.12–13 for GLM-5.1
#1
Highest-ranked open-weight model, leading by a wide margin
744B
MoE parameters, ~40B active per token (reported)
Context window: a 5× leap
The headline upgrade over GLM-5.1 — measured in tokens.
How Agent Arena scores it
No synthetic benchmarks — real, long-horizon agentic sessions scored by randomized controlled trials with live tool access.
▲ Improved vs GLM-5.1
Confirmed task success · Praise-vs-complaint
▼ Steerability slipped
−6.0% (vs +1.2% previously)
= Stable
Bash recovery · Tool hallucination
GLM-5.2 vs GLM-5.1, at a glance
Context window
1,000,000 tokens
~200K tokens
Reasoning modes
High / Max
Single mode
Launch benchmarks
None published
SWE-Bench Pro etc.
API price (per 1M)
$1.4 in / $4.4 out
Identical
MIT-licensed · weights to follow · max output 131,072 tokens · drops into Claude Code, Cline, OpenClaw.
What works
Strong coding quality, near Claude at far lower cost
Faster than predecessor; big context wins on complex tasks
Questions unclear requests — better practical steerability
Caveats
No public benchmarks at launch (SWE-Bench awaited)
Demanding tasks burn quota fast
~205 GB VRAM to run full open weights locally
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…