GLM-5.3

Z.ai released GLM-5.3, an open-source model with enhanced coding and cyber capabilities compared to its predecessor GLM-5.2, according to a benchmark analysis published in August 2026. On CyberGym, GLM-5.3 led with a score of 84.5, outperforming DeepSeek-V4Pro-0813 (83.3), Kimi K3 (80), and Qwen3.8-Max (78.5), as well as GPT-5.6 Sol (83.6) and Anthropic Fable 5 (83.8). On ExploitGym, GLM-5.3 completed 105 exploitation challenges within two hours and 130 within six hours, far surpassing GLM-5.2's 29 and 39 tasks respectively, though Anthropic's Mythos 5 remained ahead at 181 and 247 tasks. The article frames GLM-5.3 as a competitive open-weight alternative to US frontier labs, with most non-cyber benchmarks trailing top closed US models. Creator attribution remains Z.ai, not Vultr, which serves as the inference host.

Z.ai's GLM-5.3 ranks with a composite LLM Stats Score of 52.1 at a blended price of $1.33 per million tokens, per the llm-stats.com leaderboard tracking data sourced from z.ai. On CyberGym the model scores 0.84, placing sixth, and on Terminal-Bench 2.1 it scores 0.88 for seventh place, with GDPval-AA Elo of 1769/3000 for third place. The leaderboard also tracks GLM-5.3's performance across conversation depths, showing a +1.72σ improvement over seven days against the seven-day baseline as of late September 2026. Cost efficiency comparisons place it between DeepSeek-V4.1-Flash and Claude Opus 5.5 on the blended price versus score curve. The page lists FrontierSWE among additional evaluated benchmarks though scores are truncated in the supplied excerpt.