GLM-5.3 Released

• Z.ai, the international brand of Zhipu AI, put GLM-5.3 into its API, the GLM Coding Plan and the ZCode agent client on 14 August 2026 and called it its most capable model for agentic coding and defensive cyber work^[1][5][6]
• It reuses the GLM-5.2 base model untouched: Z.ai says every gain comes from a larger post-training run on harder and more varied task environments, some of which stand in for several days of an experienced engineer's work^[1][5]
• On Z.ai's own numbers it adds +50% over GLM-5.2 on its private Z.ai Code Bench, lifts Terminal Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 and Agents' Last Exam from 23.8 to 28.5, but still trails GPT-5.6 Sol on the first two (34.6 and 72.7)^[1][4][5]
• The cyber result was the surprise: 84.5% on CyberGym, ahead of the 83.8% Z.ai cites for Anthropic's restricted Mythos 5, though only 54.4% against 78.0% on ExploitBench; Z.ai says it found 2,436 vulnerabilities in 269 open-source projects with security teams in China, 1,097 of them medium or high severity^[1]
• Z.ai held the weights back for two weeks of extra safety testing, kept its most sensitive cyber functions for verified users through a "trusted access" programme and announced an "Open Source Shield" audit effort for open-source projects^[6]
• The weights arrived on 28 August under a new GLM-5.3 licence that sends companies with more than $10 billion in annual revenue through a Z.ai security review, a step away from the MIT licence of GLM-5.2^[3][4]
• On 30 September Anthropic's Frontier Red Team said GLM-5.3 completed 50 of 410 exploit attempts against 56 for Claude Mythos Preview and that its safeguards could be bypassed in 64% to 100% of simulated attempts, while NIST's CAISI called it the most cyber-capable open-weight model but well below current US frontier models^[3]

Notable features

• Text-only input with a 1M-token context window and 128K-token maximum output; thinking is always on, with low, high and max effort levels (max by default), and requests that disable thinking now fail^[1]
• About 753 billion total parameters in the published checkpoint, the same mixture-of-experts base as GLM-5.2^[4][5]
• API price of $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens, unchanged from GLM-5.2; Coding Plan access from $18 per month, with half-price points off-peak^[2][5]
• On Z.ai Code Bench at max effort it reaches 34.5% with about 75K output tokens per task against 23.4% at 96K for GLM-5.2; at high effort its 31.4% beats Claude Opus 4.8's 29.5% on about 50K tokens against 120K, while Claude Fable 5 leads at 39.5%^[1]
• Z.ai runs a public Security Disclosure Ledger that tracks each finding as it moves through disclosure^[1]

Benchmarks

Z.ai's results table from the model card, as published^[4]; the non-Z.ai columns are the maker's own comparison runs.

GDPval-AA v2 was run by Artificial Analysis, FrontierSWE by Proximal and Toolathlon Verified through the benchmark's official service^[5].