GLM-5.3 now has enough scores to rank, and it enters our main table eleventh of 681
The Aggregate Digest — Thursday, August 20, 2026
GLM-5.3 now has enough scores to rank, and it enters our main table eleventh of 681. Anyone keeping a shortlist of models worth trying has a new name for it. GLM-5.2 rates 45 ELO lower and ranks 32nd, so one version step is worth 21 places.
We called GLM-5.3 provisional on 15 August, when it had 13 scores and rated 1772. It now has 43 boards and rates 1753. Measuring it took 19 points off it, which is what the 35-board bar is for.
Z.ai published six boards with the launch and did not stack them. GLM-5.3 comes third on five of the six and first on the sixth. It wins GDPval-AA v2 at 1769 ELO against Claude Fable 5's 1743. GPT-5.6 Sol scores 34.6 on Terminal Bench 3.0 to GLM-5.3's 28.3, and gets 293 exploits to its 130 on ExploitGym over six hours. GLM-5.2 finishes last on four of the six.
The caveat is a tie. GLM-5.3 and Gemini 3.6 Flash both rate 1753, and the order between them is a tiebreak. We have measured GLM-5.3 on 43 boards and Gemini 3.6 Flash on 182.
No crown moved today. Gemini 3.7 Flash has been the best available model since 15 August and rates 1787, which puts GLM-5.3 34 points below it.
