GLM-5.3 Gets 50% Better Coding Without Bigger Model — Challenges Frontier AI

Z.ai's new GLM-5.3 is making a serious push into frontier-model territory, particularly in coding, agentic tasks, and cybersecurity.

Key points

• Terminal-Bench 2.1 88.2 88.3 87.9 86.6 85.0 88.0 88.8.
• DeepSWE v1.1 66.9 67.5 62.7 56.6 58.0 69.7 72.7.
• At Max effort, GLM-5.3 reached 34.5% at roughly 75, 000 output tokens, compared with GLM-5.2's 23.4% at around 96, 000.
• Benchmark GLM-5.3 Kimi K3 DeepSeek V4 Pro Qwen3.8 Max Opus 4.8 Fable 5 GPT-5.6 Sol.
• Terminal-Bench 3.0 28.3 17.4 – – 21.1 33.7 34.6.

For context, the interesting part is that Z.ai did not create the model bigger. GLM-5.3 uses the same base model as GLM-5.2, with its improvements coming from expanded post-training on more environments, more diverse tasks, and longer-running engineering workloads. Z.ai notes this resulted in a 50% improvement on its internal coding benchmark. GLM-5.3 Tops KingBench 3.

Meanwhile, the headline result comes from KingBench 3, a third-party, creator-run benchmark by AICodeKing that uses the same eight coding and simulation tasks for each model.

GLM-5.3 scored 73 out of 80, or 91.25%, compared with 75% for GLM-5.2 around two months earlier. Model KingBench 3 Score. GLM-5.3 91.25% (73/80). Fable 5 82.5%. Qwen3.8 Max 81.25%. Opus 4.8 80%. Opus 5 77.5%. Kimi K3 77.5%.

That puts GLM-5.3 comfortably ahead of the other models tested on this particular suite.

Nevertheless, KingBench 3 should not be confused with a substantial independently governed benchmark such as Terminal-Bench. It is a relatively small eight-task benchmark created by an independent tester rather than Z.ai, so the result is useful evidence but not proof that GLM-5.3 is universally better than every frontier model. Users Claim OpenAI Secretly Cut Codex Usage Limits by Half. Z.ai's broader results provide a more balanced comparison.

DeepSWE v1.1 66.9 67.5 62.7 56.6 58.0 69.7 72.7. NL2Repo 58.0 58.0 61.1 55.9 69.7.

ProgramBench 19.0 17.5 – 10.5 15.5 33.0 23.0. FrontierSWE 78.1 – – – 66.5 88.2.

Meanwhile, the results show that GLM-5.3 is clearly competitive, but it does not dominate every benchmark.

For example, it almost matches GPT-5.6 Sol, Kimi K3, and Fable 5 on Terminal-Bench 2.1. On the much harder Terminal-Bench 3.0, GLM-5.3 scores 28.3, well above GLM-5.2's 4.6 and Opus 4.8's 21.1, but still behind Fable 5 at 33.7 and GPT-5.6 Sol at 34.6.

That jump from 4.6 to 28.3 in only one generation is arguably more significant than GLM-5.3 winning any single benchmark. Where GLM-5.3 Improved Most.

KingBench additionally highlights how much better GLM-5.3 has become at combining working logic with polished interfaces. KingBench 3 Task GLM-5.3 Score. Elevator simulation 8/10. 3D contact lens case 8/10. 3D folding table 10/10. Panda eating burger SVG 10/10. Bow and arrow game 10/10. Permutation difficulty 10/10. Gemma 2B fine-tuning pipeline 10/10. Dual-time-zone 3D watch 7/10.

In practice, the 3D wristwatch was particularly tough. While Fable 5 scored 4 and Opus 5 scored 3, GLM-5.3 scored 7/10. The generated watch included moving real-time hands, a smooth second hand, day and date displays, and a second time zone.

GLM-5.3 additionally improved dramatically on the contact-lens task, rising from 3/10 for GLM-5.2 to 8/10 despite using the same base model. GLM-5.3 Is Particularly Solid in Cybersecurity.

Coding is only part of Z.ai's focus. GLM-5.3 has been heavily post-trained for code auditing, vulnerability discovery, and cyber defense.

Cyber/Agent Benchmark GLM-5.3 GLM-5.2 Fable 5 GPT-5.6 Sol. CyberGym 84.5 77.2 83.8 83.6. ExploitBench 54.4 24.4 78.0 76.5. AutomationBench 48.2 26.2 46.2 45.8. Agents' Last Exam 28.5 23.8 23.8 28.6.

This reveals an significant distinction. GLM-5.3 is already highly competitive at finding vulnerabilities, but frontier closed models still have a substantial advantage when tasks move deeper into actual exploitation.

On ExploitGym, GLM-5.3 completed 105 tasks within a normalized two-hour budget and 130 within six hours. While GPT-5.6 Sol reached 216 and 293, fable 5 completed 181 and 247 respectively.

Z.ai notes its security work has already identified 2, 436 vulnerabilities throughout 269 real-world projects, including 1, 097 critical or high-severity findings. Those are Z.ai's own documented figures and remain part of its ongoing disclosure program. ChatGPT's New Memory Records Your Clicks and Typing in Plain Text. Post-Training is the Biggest Story.

Perhaps the most significant part of GLM-5.3 is not that it beats another model on one leaderboard.

Z.ai notes the base model did not change from GLM-5.2. Instead, it scaled post-training using more environments, harder long-horizon tasks, reinforcement learning, and extra compute.

That additionally appears to have improved efficiency. On Z.ai's private Code Bench, GLM-5.3 reached a 31.4% completion rate at High effort while producing roughly 50, 000 output tokens per task. Opus 4.8 scored 29.5% while using around 120, 000 tokens.

At Max effort, GLM-5.3 reached 34.5% at roughly 75, 000 output tokens, compared with GLM-5.2's 23.4% at around 96, 000. Fable 5 remained ahead overall at 39.5%.

These are Z.ai's internal results, so they cannot be treated the same way as public benchmark scores. Still, they suggest GLM-5.3 may be doing more work with fewer generated tokens. Is GLM-5.3 Better Than Frontier Models? Not throughout the board.

GLM-5.3 at present looks strongest when the workload involves coding agents, simulations, frontend plus backend development, automation, and defensive security.

It beats a number of frontier models on KingBench 3 and CyberGym and comes surprisingly close on Terminal-Bench 2.1. But GPT-5.6 Sol and Fable 5 remain stronger on a number of challenging long-horizon coding and exploitation benchmarks.

For context, the broader takeaway is that GLM-5.3 has moved much closer to the closed frontier without increasing the size of its underlying model. For developers, that may be more significant than winning any single benchmark: the gap between leading open and closed coding models is becoming increasingly narrow. Stay Connected with ProPakistani.

Obtain the latest tech news, telecom insights, and product launches wherever you prefer. Follow on Google Discover.

Add as a preferredSource on Google Follow on Google News Join WhatsApp.

Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.

Technology and Automotive Specialist covering the latest cars, smartphones, AI breakthroughs, and.

For now, GLM-5.3 Gets 50% Better Coding Without Bigger Model remains the part of the story worth watching, and further updates are likely as more details are confirmed.