Chinese startup Z.ai has unveiled GLM-5.3, securing a score of 60 on the Artificial Analysis Intelligence Index. That result puts the model in a dead heat with Kimi K3 for the lead among open architectures, advancing seven points beyond its predecessor, GLM-5.2.
The defining gains land squarely in autonomous workflows. On the agentic GDPval-AA v2 benchmark, GLM-5.3 jumped 246 Elo points to reach 1,770 Elo—settling into second place overall, trailing only Anthropic's Claude Opus 5 at 1,855. On the economics side, inference runs at $0.68 per task according to Artificial Analysis. While that marks a cost increase over GLM-5.2's $0.44 footprint, it undercuts Kimi K3's $0.84 price tag by 19 percent, stepping up commercial pressure on Western proprietary APIs.
Yet the distribution strategy reveals a familiar playbook. Despite being branded as an open-weights milestone, Z.ai has gated GLM-5.3 behind its proprietary API, postponing public weights for roughly two weeks under the premise of auditing automated vulnerability discovery. In practice, this gives the startup an exclusive window to monetize commercial inference traffic before releasing open checkpoints to the developer ecosystem.