Skip to content

Model Card|GLM-5.3

Aug 25, 2026 1 min
TL;DR GLM-5.3: same GLM-5.2 base model, pure post-training gains, 1M context / 128K max output, pricing unchanged at $1.4 input / $4.4 output (per 1M tokens), Terminal-Bench 3.0 jumps from 4.6% to 28.3% (open-source SOTA), CyberGym vulnerability discovery 84.5% surpasses every listed closed-source frontier model. Zhipu delayed weight release until safety evaluation completes (~8/28)
Table of Contents
  1. Model Information
  2. Key Capabilities
  3. Benchmark Results
  4. Comparison with Predecessor and Competitors
  5. What This Means for Agent Development
  6. Takeaway
  7. References

🌏 中文版

Model Information

FieldValue
Model IDglm-5.3
VendorZhipu AI (international brand: Z.ai)
ParametersNot officially disclosed; shares GLM-5.2 base model (third-party estimates ~744B total / ~40B active parameters)
Context Window1,000,000 tokens (max output 128,000 tokens)
Input pricing (USD/1M tokens)$1.40 (cached input $0.26)
Output pricing (USD/1M tokens)$4.40
Open-sourceNot yet — Zhipu announced weight release within two weeks of launch (~2026-08-28), pending safety evaluation and hardening
Release date2026-08-14
Official announcementZ.ai Blog: GLM-5.3
HuggingFacezai-org/GLM-5.3 (page shows Coming Soon, not yet downloadable)
FamilyGLM 5.x (shares the same base model as GLM-5.2)

Key Capabilities

  • Terminal-Bench 3.0 jumps from GLM-5.2's 4.6% to 28.3%, achieving open-source model SOTA
  • Z.ai Code Bench (internal benchmark) improves 50% over GLM-5.2; Zhipu calls it "the strongest open-weight code model to date"
  • CyberGym vulnerability discovery score of 84.5% — the highest among all models in the official comparison table, including closed-source frontier models
  • Collaborated with multiple security teams on real-world targets, uncovering 2,436 vulnerabilities, of which 1,097 were medium-to-high severity
  • No base model retraining whatsoever — all improvements come from post-training scaling (more task environments + more compute)

Benchmark Results

BenchmarkGLM-5.3GLM-5.2Best competitor
Terminal-Bench 3.028.3%4.6%GPT-5.6 Sol 34.6%
CyberGym (vulnerability discovery)84.5%77.2%Claude Fable 5 83.8%
ExploitBench (exploitation)54.4%24.4%Claude Fable 5 78.0%
DeepSWE v1.166.9%46.2%GPT-5.6 Sol 72.7%

⚠️ All results above are self-reported by Z.ai; independent third-party reproductions are not yet available.

Comparison with Predecessor and Competitors

GLM-5.3 shares the exact same base model as GLM-5.2 — there was zero pre-training this time around. Every improvement comes from post-training scaling, which is literally the first sentence of the official blog post: "Scaling post-training is all we did for GLM-5.3." Terminal-Bench 3.0 jumped from 4.6% to 28.3%, a 6x leap. While it still trails GPT-5.6 Sol's 34.6% and Claude Fable 5's 33.7%, it is now the highest-scoring open-source model.

The CyberGym result stands out: GLM-5.3's 84.5% not only beats GLM-5.2's 77.2%, it also surpasses every closed-source frontier model in the official comparison table, including Claude Fable 5 (83.8%) and GPT-5.6 Sol (83.6%). This is a rare case of an open-source model overtaking closed-source frontier models in security vulnerability discovery. ExploitBench (one step deeper into the exploitation chain) also doubled, though the absolute score (54.4%) still clearly trails Claude Fable 5 (78.0%).

Pricing stays completely unchanged at GLM-5.2's $1.4 / $4.4 (per 1M tokens) — with substantially improved capabilities, this amounts to an implicit price cut. However, weight release has been delayed until "safety evaluation and hardening are complete," a departure from GLM-5.2's day-one open-source MIT release. Zhipu is effectively acknowledging through action that the leap in cybersecurity capabilities brings new risk assessment obligations.

What This Means for Agent Development

The CyberGym / ExploitBench scores are the most noteworthy signal here: an open-source model's vulnerability discovery and exploitation capabilities have caught up to — and in some cases surpassed — closed-source frontier models, meaning the barrier to automated security testing (or conversely, automated attacks) is dropping fast.

  • If you're building coding agents or automated PR review: The open-source SOTA score on Terminal-Bench 3.0 means local deployment (once weights are released) can deliver solid terminal operation and long-horizon task capabilities — worth adding to your self-hosting evaluation list
  • If you're building DevSecOps or automated vulnerability scanning tools: GLM-5.3's CyberGym / ExploitBench performance is best-in-class, making it a strong fit for white-box code auditing and automated discovery of known vulnerability patterns — but plan access controls in parallel, to prevent the same capabilities from being used offensively
  • Not suitable for: There is no public API or downloadable weights yet — access is only through the GLM Coding Plan subscription. If you need immediate offline inference or custom fine-tuning, you'll need to wait roughly one week (~8/28) for the weight release

Takeaway

The conventional wisdom was that open-source models needed bigger base models or full retraining to catch up to frontier capabilities. GLM-5.3 used the same GLM-5.2 base model, relying purely on post-training scaling, to boost Terminal-Bench 3.0 by 6x and even overtake every closed-source frontier model in security vulnerability discovery. This suggests that "post-training craft" — the scale and sophistication of the post-training pipeline and task environments — is becoming a more valuable moat than pre-training scale itself. And the cybersecurity spillover from this approach has already forced the vendor to delay its own open-source timeline for risk assessment.

References