GLM-5.3 Surges 246 Points, Ranks Second to Claude

Why this matters A 246-point Elo jump on GDPval-AA v2 is the kind of number that would normally headline itself, but the fine print matters more here. Z.ai is closing the agentic-task gap with Claude Opus 5 while still pricing GLM-5.3 well below it, at $0.68 versus whatever premium Anthropic charges, which is the real pitch to developers weighing frontier capability against burn rate. The catch: that price tag is 1.5 times what GLM-5.2 cost, so "cheap" is relative, and the delayed release means none of this is testable in production yet.

For founders building agent pipelines on open models, the tie with Kimi K3 at the top of the open-model index is worth watching, but benchmark Elo and real-world reliability on messy, multi-step tasks are different animals. We'd treat this as a strong signal that Chinese labs are optimizing specifically for agentic workflows rather than general chat quality, and that's the metric to track once GLM-5.3 actually ships and outside teams can run it against their own task suites.