GLM 5.3 adds no parameters, and its sharpest jump is in finding vulnerabilities

Zhipu AI released GLM 5.3 on August 14, 2026, and opened the announcement with a one-sentence thesis: "Scaling post-training is all we did for GLM-5.3." ^[1][2] The model is the same roughly 744-billion-parameter Mixture-of-Experts brain as GLM 5.2, with about 40B parameters active per token, which means the entire claimed capability jump has to be paid for by training method rather than by size. ^[3][4] That is about 5.4 percent of weights firing per token, unchanged. ^[5]

The training stack is also unchanged: IndexShare for long-context efficiency, SAO for reinforcement learning on long-horizon tasks, and slime, the open-source RL framework from THUDM. ^[9] What scaled was the environment supply. Zhipu says research agents mine patterns from real engineering work and compile them into runnable multi-step tasks with hidden state, a judge agent attempts each one to confirm it is solvable, verifiers are synthesized without sight of the reference solution, and solver trajectories are mined to find and close reward shortcuts. ^[10] Some environments represent several days of work for an experienced engineer, such as diagnosing bottlenecks on a training stack with access to clusters, storage, codebases, docs and experiment results. ^[11] The enabling number is a systems number: workload-aware scheduling and related optimizations raised end-to-end RL training throughput by more than 2.3x on long-horizon coding tasks. ^[12]

On coding, the vendor-reported deltas are large and the token counts are the interesting part. On Zhipu's private Z.ai Code Bench at Max effort, GLM 5.3 scored 34.5 percent using roughly 75K output tokens per task against GLM 5.2's 23.4 percent at 96K ^[13], which is about 47 percent more score on 22 percent fewer tokens ^[27]. At High effort it reported 31.4 percent at about 50K tokens versus Claude Opus 4.8's 29.5 percent at 120K, roughly 58 percent fewer tokens for a higher score. ^[14][28] It still trails Fable 5 at 39.5 percent. ^[15]

Public benchmarks, all vendor-published: Terminal-Bench 3.0 goes to 28.3 from 4.6, roughly a 6x move, against Kimi K3 at 17.4 and closed leaders at 33.7 and 34.6 ^[17][18]; SWE-Marathon v1.1 to 42.5 from 19.4 ^[23][29]; DeepSWE v1.1 to 66.9 from 46.2 ^[20]; FrontierSWE to 78.1 from 67.5 ^[22]; Agents' Last Exam (CLI) to 28.5 from 23.8 ^[21]; PostTrainBench to 39.8 from 31.7 ^[24]; Terminal-Bench 2.1 to 88.2 from 81.0 ^[19]. GLM 5.3 leads the open-weight class on four of those and Kimi K3 still holds three. ^[25]

The security claim is the one that should get scrutiny, because it is the biggest and the thinnest. Zhipu says "emergent cyber capability" arrived faster than the team expected, that GLM 5.3 tops CyberGym, a vulnerability-discovery benchmark, and that it more than doubles GLM 5.2 on exploitation benchmarks. ^[7] No score is attached to either statement in the material. Housekeeping supports the impression of a fast assembly: the same Anthropic scores appear as "Fable 5 (w/ fallback)" in the benchmark table and as "Mythos 5" in the cyber section. ^[16] Hacker News readers also flagged omissions from the comparison table. ^[26]

Nothing here is independently verified. The standalone API is listed as coming soon and the weights are about two weeks out. ^[8] Watch for a CyberGym reproduction by anyone outside Zhipu once weights land, whether the Kimi K3 leads on Terminal-Bench 2.1, DeepSWE and SWE-Marathon flip, and whether the token-efficiency gain survives on public tasks rather than the private bench.