AI Pulse · AI Trends Pulse

Z.ai, the company behind the GLM model family, is claiming a big jump in coding ability with its new GLM-5.3 model. According to the discussion on Hacker News, the company says GLM-5.3 beats its own predecessor, GLM-5.2, by 50% on coding capability, and that gain came purely from post-training work, not from a bigger or different base model. That’s worth noting because it means labs are still finding real gains by refining models after the initial training run, not just by throwing more compute at bigger models.

The other claim getting attention is on security. Z.ai says GLM-5.3 edges out Anthropic’s Mythos and OpenAI’s GPT-5.6 Sol on a benchmark called CyberGym, which tests how well a model handles cybersecurity tasks. The scores are close, 84.5% for GLM-5.3 against 83.8% and 83.6% for the other two. That’s a narrow lead, not a blowout, and these are self-reported numbers from the company, so treat them as a claim rather than an independently confirmed result.

For business owners, the practical point is this. If GLM-5.3 holds up once people can actually test it, it adds another strong, cheaper option to the coding and security tooling market, which keeps pressure on pricing across the board for anyone buying AI capability. Nothing is public yet though. Weights aren’t released, and Z.ai says it needs about two weeks for safety hardening before that happens. So there’s nothing to plug in today. Worth watching once real users get their hands on it and start comparing it against what you’re already using. This is the kind of competitive shift we track for clients inside an AI command centre, so tools that actually move the needle get flagged before the hype settles.