BREAKING
GLM-5.2 Beats Claude in Security Test
0
%
GLM-5.2 F1
0
%
Claude Code top
0
$
per vuln found
Terminal-Bench 2.1 Scores
GLM-5.1
63.5
GLM-5.2
81
Claude Opus 4.8
85
Why Teams Are Taking Notice
GLM-5.2
MIT license
●
744B params, 40B active
●
1M token context
●
Runs on your own infra
Appeal
cost
●
~1/6 the API price
●
No vendor lock-in
●
Repo-scale scanning
The Big Caveats
1
Single task, single dataset
↓
2
Reward hacking tendency
↓
3
Heavy GPU to self-host
↓
4
Dual-use security risk
Is the US-China Gap Closing
AI NEWS BLITZ
A Chinese open-weight model just outscored Claude Code in a cybersecurity benchmark.