大模型网摘GLM-5.3 Model Deep Dive What the Post-Training Upgrade Delivers
The release rhythm first
Zhipu shipped GLM-5.3 on August 14, 2026. There was a prelude. The model page briefly went public for about an hour on August 3, enough for developers to skim the docs, and the Chinese community treated the slip as good gossip. On launch day GLM Coding Plan quotas were reset for all users, with an official tech blog to match.
Same architecture, gains from post-training
According to Yicai, GLM-5.3 keeps 753 billion total parameters, essentially unchanged from GLM-5.2, and the base model was not swapped. The gains come from longer post-training. Official numbers put internal coding benchmarks about 50 percent above GLM-5.2, with coding and agent capability described as close to Claude Fable 5. These claims rest on vendor benchmarks for now. Independent retests need time.
Security is the other focus. Multiple outlets noted targeted reinforcement on code vulnerability discovery and related security tasks, which lines up with rising enterprise attention to model safety in procurement.
Pricing and open weights
Third-party summaries of the official price page list GLM-5.3 at 1.4 US dollars input and 4.4 dollars output per million tokens, with cache reads at 0.26 dollars. Identical to GLM-5.2. An upgrade without a price hike is uncommon among this year's flagships. Weights are released openly for self-hosting. For context, our October open-weight flagship spec sheet places GLM-5.3 alongside Kimi K3.1 Pro, DeepSeek V4-Pro-0813 and Qwen3-Max.
GLM-5.3-Flash drives the price down
On August 26 Zhipu followed up with GLM-5.3-Flash. 320 billion total parameters, 18 billion active, native multimodal input new to the GLM-5 series, and a 1M token context. Before launch it was blind-tested at scale under the anonymous name Ox-Alpha on OpenCode and OpenRouter. Chinese users nicknamed it Niulai. Third-party evaluations put its Artificial Analysis intelligence index at 57, on par with Claude Opus 4.8, whose official API pricing is 5 dollars input and 25 dollars output per million tokens. Media coverage described the price gap as roughly one fortieth. The Flash weights are open-sourced under MIT on HuggingFace.
Ecosystem and tooling
The coding toolchain keeps moving. Zhipu's own coding agent ZCode released version 3.17.2 on November 3, aimed mainly at GLM Coding Plan subscribers. Third-party comparisons have appeared too. Our DeepSeek Flash vs GPT-6.1 Sol and GLM-5.3 Flash piece reads this batch of models from a selection standpoint.
What to watch
Two things merit attention. Independent retests of the GLM-5.3 coding claims, and longer-run community feedback on real engineering work. Our earlier piece on the claim that open models handle most agentic work at a fraction of the cost applies to judging GLM-5.3's real-world output as well. This article is compiled from publicly available online sources.