GLM-5.3

GLM-5.3 is Z.ai's August 2026 flagship, and the interesting part is how it was built: not a new pretrain, but extended post-training on the same base as GLM-5.2. The gains are concentrated where that kind of work pays off. Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE v1.1 went from 46.2 to 66.9, and independent evaluation puts it at 95.4 on SWE-bench Verified. Z.ai also reports 62.5 on Humanity's Last Exam with tools and a notable cybersecurity capability that emerged rather than being targeted, with 54.4 on ExploitBench and 84.5 on CyberGym. Pricing is $1.40 per million input tokens and $4.40 output, with a 1M token context window and 128K max output. It launched API-first through the Z.ai API and Coding Plan; the Hugging Face repository is a gated placeholder counting down to August 28, 2026, so the weights are announced rather than available. Z.ai has not published a parameter count or a license for this release, and its predecessor's plain MIT terms should not be assumed to carry over. Treat every score here as vendor-reported except the independently run SWE-bench figure.