GLM 5.3 vs GLM 5.2: What's Actually New (Benchmarks Included)

Quick answer: GLM 5.3 and GLM 5.2 share the same base model — Z.AI improved 5.3 purely through scaled post-training. The practical deltas: roughly +50% coding capability on Z.AI's internal code bench, open-weights SOTA on Terminal-Bench 3.0 (28.3 vs 4.6) and Agents' Last Exam, more than double the exploitation capability on ExploitBench, and a mandatory API change (thinking can no longer be disabled).

TL;DR

The One-Line Summary

GLM 5.3 is GLM 5.2 trained harder — not a new model. Same pretrained base, same architecture, same context behavior. Everything that changed came from one month of scaling post-training: more long-horizon environments, more diverse tasks, and more RL compute. If you're asking "is GLM 5.3 worth it over 5.2?", the answer is "yes, if you do agentic coding or security work — and the migration cost is one JSON field."

Coding: Where the Gap Is Real

The coding deltas are the reason to upgrade. Z.AI reports a 50% improvement over GLM 5.2 on its in-house Z.ai Code Bench, and the public benchmarks agree:

Terminal-Bench 3.0 is the big one. GLM 5.2 scored 4.6 — near-baseline on the harder benchmark — and GLM 5.3 jumps to 28.3, now in range of closed frontier models (GPT-5.6 Sol: 34.6, Fable 5: 33.7, Opus 4.8: 21.1). The same pattern holds on DeepSWE and SWE-Marathon: the harder the benchmark, the larger the improvement — exactly what you'd expect from RL that specifically targets long-horizon task completion.

On Z.ai Code Bench (Z.AI's private, contamination-resistant benchmark of realistic dev environments), GLM 5.3 wins on both axes:

• Max effort: 34.5% task completion at ~75K output tokens — vs GLM 5.2's 23.4% at ~96K tokens.
• High effort: 31.4% at ~50K tokens, beating Claude Opus 4.8 (29.5% at ~120K) while using 2.4× fewer tokens.

Better completion and better token efficiency is the signature of post-training that actually works.

Agents: Longer Horizons, More Reliability

AutomationBench nearly doubled (26.2 → 48.2). This is the benchmark that measures autonomous business-process automation — long sequences of API calls, form fills, and state checks — and it's the strongest signal that 5.3 keeps its head on long agent runs where 5.2 dropped steps.

Cyber: The Emergent, and Controversial, Delta

Z.AI is unusually candid here: they added vulnerability-discovery data to the training mix and the capability "developed faster than we expected."

CyberGym's 84.5 is the best public result on the benchmark, ahead of Mythos 5 (83.8) and GPT-5.6 Sol (83.6). ExploitBench more than doubled — and Z.AI notes the gains are largest further up the exploitation chain, i.e., exactly where the dual-use risk concentrates. That's also why weights are held for two weeks after launch while safety evaluation and hardening finish. If you use GLM for security work, this is a step-change; if you deploy it, budget for the policy conversation.

API Migration: One Required Change

GLM 5.3 removes support for disabling thinking. If your code sends thinking.type: "disabled", requests will fail after you switch the model ID.

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "max"
}

• reasoning_effort: low | high | max (default max; Z.AI recommends max for coding).
• Migration path: set enabled + low before updating the model ID, then tune the effort level.

Everything else — endpoint, context, streaming, structured output — carries over from 5.2.

Pricing & Access

• GLM Coding Plan: all subscribers were moved to GLM 5.3 on launch day. The plan is now points-based (input / cached input / output metered separately); off-peak hours (outside 14:00–18:00 UTC+8, Mon–Fri) cost 50% of standard points.
• ZCode: 98%+ cache hit rate, 1.5× limited-time quota boost through August 31, Goal mode for long-horizon tasks, Remote Control from WeChat/Feishu.
• API: model id glm-5.3, per-token billing.
• Open weights: HuggingFace release in two weeks (safety permitting).

Which Should You Use?

Upgrade to GLM 5.3 if you: do agentic coding (multi-file changes, long task runs), run automated workflows, care about token cost per completed task, or work in security/vulnerability research.

Stick with GLM 5.2 if you: need to keep thinking disabled (only 5.2 supports it), have locked-in prompts that can't tolerate the effort-level behavior change, or prefer the stable, already-hardened weight release.

For everyone else, the migration is one field and the benchmarks are unambiguous — GLM 5.3 is the better model.

FAQ

Is GLM 5.3 better than GLM 5.2 for coding? Yes. ~50% better on Z.AI's internal code bench; big jumps on Terminal-Bench 3.0 (4.6→28.3), DeepSWE (46.2→66.9), and SWE-Marathon (19.4→42.5).

Is GLM 5.3 a new architecture? No. Same base model as GLM 5.2; all gains are from scaled post-training.

Can I still disable thinking in GLM 5.3? No. thinking.type: "disabled" is no longer supported — use enabled with reasoning_effort: "low" for the lightest mode.

When do GLM 5.3 weights come out? Two weeks after the August 14 launch, after safety evaluation and hardening.

Is GLM 5.3 more expensive than GLM 5.2? Same per-token API pricing model. Coding Plan is now points-based, with 50% off-peak pricing.

Is GLM 5.3 safe to deploy? It's a dual-use concern — cyber capability more than doubled — which is why Z.AI holds weights two weeks for safety work. API access is available now.

Sources

• Z.AI: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (August 14, 2026)

Last updated: August 14, 2026