A model release is one thing. The verdict of the people who work with these models every day is another — and this week it’s surprisingly clear. Nathan Lambert (interconnects.ai) just published an analysis whose title says it all: GLM-5.2 is the ‘step change’ for open agents.
What this is about
GLM-5.2 from Chinese lab Z.ai has been available as an open-weight model under an MIT license for a little over a week — 753 billion parameters, only about 40 billion of them active. We already covered the model itself. What’s new is the assessment: in community benchmarks it punches above expectations, and on Arena’s agent leaderboard it’s, per Lambert, the only open model keeping pace with OpenAI’s and Anthropic’s current top models.
Simon Willison had already called it “probably the most powerful text-only open weights LLM.” So two of the most level-headed voices in the space land on a similar conclusion, independently.
Why this is more than a benchmark win
The point isn’t that one model gains a few percentage points. The point is the shift behind it. On agentic tasks — multi-step planning, tool use, staying coherent over long runs — the gap between open and closed models was long considered wide. That’s exactly where it’s now visibly closing. If you want to self-host an open model, fine-tune it, or simply stay independent of any single vendor, you suddenly have a serious option.
My take
For us Claude users this isn’t bad news — quite the opposite. Pressure from the open side keeps everyone honest and forces all of them to get better. Anthropic and OpenAI still lead at the very top, but the comfortable margin is shrinking.
I’ll stay realistic, though: “keeps pace” doesn’t mean “just as good” in daily use, and running a 753B model locally is no weekend project. But the direction is unmistakable. If open agents keep this up, the interesting question won’t only be which model is best — it’ll be who owns the infrastructure we build on.
Sources: interconnects.ai: GLM-5.2 is the step change for open agents, Simon Willison: GLM-5.2