Zhipu AI's GLM-5.2 beat Anthropic's Claude Opus 4.8 on Semgrep's cybersecurity bug-hunting benchmark. With additional instructions, the Chinese model matched the US frontier model's performance, narrowing the gap.
Key facts
- Zhipu GLM-5.2 released June 13, 2026
- Beat Anthropic Claude Opus 4.8 on Semgrep bug-hunt benchmark
- Matched US model with additional instructions from Semgrep
- Hailed as another 'DeepSeek moment' for Chinese AI
- DeepSeek raised $7B on June 27, 2026
Beijing-based start-up Zhipu AI released GLM-5.2 on June 13. In testing by cybersecurity firm Semgrep, the model outperformed Anthropic's Claude Opus 4.8 on bug-hunting tasks, according to The Wall Street Journal. When Semgrep researchers provided further instructions, GLM-5.2 matched the US model's performance entirely.
Another 'DeepSeek moment'
The result has been hailed as another 'DeepSeek moment' for Chinese AI, echoing how DeepSeek's V4-pro and R1 models matched frontier US models at a fraction of the training cost. Zhipu AI, which raised significant funding in 2025, is now demonstrating competitive capability in a high-stakes domain: cybersecurity. The benchmark tests show Chinese models are closing the gap not just on general reasoning but on specialized, safety-critical tasks.
The broader context
This comes as China's AI labs accelerate open-source releases. Earlier this week, Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter model trained entirely on domestic chips. DeepSeek, meanwhile, raised $7 billion in its first major funding round on June 27, abandoning its no-funding pledge. The competitive pressure on US frontier labs like Anthropic is mounting from multiple Chinese players simultaneously.
Anthropic has been under regulatory scrutiny recently: it voluntarily suspended Claude Mythos on June 26 under regulatory pressure, and is targeting an IPO at a $1 trillion+ valuation. The company's Claude Opus 4.8 is the latest iteration of its flagship model, scoring 88.6% on SWE-bench Verified and 78.9% on Terminal-Bench 2.1. That Zhipu's GLM-5.2 can beat it on a cybersecurity benchmark is a concrete data point, not just a narrative.
What to watch
Watch for Anthropic's response: an updated Claude Opus release or a new cybersecurity-focused benchmark. Also track whether Zhipu AI open-sources GLM-5.2, as DeepSeek did with V4-pro, and whether Semgrep releases the full benchmark methodology publicly.
Source: scmp.com
[Updated 03 Jul via scmp_tech]
Zhipu AI on Wednesday released ZCode, a harness for GLM-5.2 designed to help developers build autonomous coding assistants, intensifying its rivalry with Anthropic [per SCMP]. The move comes as Anthropic removes hidden code tracking Chinese users of its Claude Code platform, and follows Alibaba's ban of Claude Code over alleged backdoor risks. Zhipu, listed in Hong Kong as Knowledge Atlas Technology, is among multiple Chinese firms—including Tencent, Alibaba, and ByteDance—launching competing AI coding agents that leverage localized workflows to challenge US models.
