Zhipu Releases GLM-5.3, Edging OpenAI and Anthropic on CyberGym

Beijing-based AI lab Zhipu, also known as Z.ai, has shipped GLM-5.3, its newest flagship coding model, and claims it now tops the CyberGym cybersecurity benchmark — a narrow but symbolically significant edge over Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol.

What happened

GLM-5.3 reuses the same roughly 700-billion-parameter base model as GLM-5.2, the version Zhipu shipped in June. Every capability gain comes from extended post-training rather than a new architecture, according to The Decoder. Zhipu reports a roughly 50% jump in coding capability and calls GLM-5.3 the strongest open-weights coding model available, with the largest gains in agent-based tasks.

The headline number is cybersecurity. On CyberGym, a benchmark that measures whether models can identify and validate security flaws in source code, Zhipu reports GLM-5.3 reached 84.5% — just ahead of Anthropic's Mythos 5 at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%, per the South China Morning Post.

The lead does not extend everywhere. On ExploitBench, which gauges how far a model climbs a full exploitation chain, GLM-5.3's 54.4% trails Mythos 5's 78% and GPT-5.6 Sol's 76.5%. The pattern is consistent: GLM-5.3 is now competitive at finding vulnerabilities, but US frontier models still go deeper once an exploit is underway.

Zhipu also tested the model with security teams in China against real codebases, reporting 2,436 vulnerabilities found across 269 projects — 1,097 of them medium or high severity, including flaws in projects up to 40 years old.

Why it matters

The weights go open source in two weeks under a permissive license, once security reviews wrap up. GLM-5.3 is already live in the GLM Coding Plan and works with agent harnesses including ZCode, Claude Code, and OpenCode.

For AI builders, the release matters on three fronts. First, it is the clearest signal yet that Chinese open-weight models have closed the coding gap with closed frontier systems — and are now competitive in security-specific benchmarks. Second, the two-week open-source window gives teams a concrete path to self-host a frontier-class coding model. Third, the CyberGym result signals that vulnerability discovery is becoming a mainstream model capability rather than a specialist niche.

The launch lands as Chinese labs — Moonshot AI's Kimi, DeepSeek's V4 models, and Alibaba's Qwen — keep narrowing the performance gap with US leaders while undercutting them on price and openness. GLM-5.3 is the latest evidence that frontier capability is no longer a two-company race.