Their new open-source GLM-5.1 beat closed-source flagships like GPT-5.4 and Claude Opus 4.6 on the hardcore SWE-Bench Pro eval (more charts here, you can try it here). The model can autonomously dig through repositories for up to 8 hours, running thousands of iterations until it ships a working fix.
Their researchers' pace is impressive, but the economics still have to add up somehow. Bloomberg reports that Z.ai's revenue has dipped (unlike, say, MiniMax), so the era of price dumping in Chinese AI seems to be winding down. The team raised API pricing for GLM-5.1 by at least 8%, with Alibaba (which has its own capitalization issues) and Tencent quickly following suit. It looks like the "capture the market at any cost" phase is finally giving way to actual monetization attempts.