TLDR Z.ai released GLM-5.3 on August 14, 2026, on the exact same ~743–753B-parameter Mixture-of-Experts base model as GLM-5.2. Every gain came from scaled post-training on long-horizon environments. The result is a ~50% relative improvement on Z.ai’s internal Code Bench, open-source SOTA on Terminal-Bench 3.0 (4.6 → 28.3) and Agents’ Last Exam, dramatically better agentic scores, and emergent state-of-the-art cybersecurity performance (CyberGym 84.5%). Token efficiency also improved.
Weights are expected roughly two weeks after launch after safety hardening. Developers can already access related GLM models and test workflows efficiently through unified platforms such as CometAPI.
Key Takeaways
- Same base model as GLM-5.2; all advances from post-training (IndexShare, SAO, slime framework + more environments and compute).
- Coding: Terminal-Bench 3.0 4.6 → 28.3; DeepSWE v1.1 46.2 → 66.9; internal Z.ai Code Bench ~50% better with fewer output tokens.
- Agentic: AutomationBench 26.2 → 48.2; Agents’ Last Exam 23.8 → 28.5; GDPval-AA v2 1508 → 1769.
- Cyber: CyberGym 77.2 → 84.5 (SOTA, ahead of Mythos 5 and GPT-5.6 Sol); ExploitBench more than doubled (24.4 → 54.4). Real-world finds: 2,436 vulnerabilities across 269 projects.
- Specs unchanged in core architecture: 1M context, up to 128K output tokens, always-on thinking with low/high/max effort.
- Access: Live via GLM Coding Plan and ZCode; general API and open weights (expected MIT-style) coming after safety review. CometAPI offers convenient OpenAI-compatible access to the GLM family for side-by-side testing and production routing.
GLM-5.3 vs GLM-5.2 at a Glance
| Dimension | GLM-5.3 | GLM-5.2 | Winner / Notes |
|---|---|---|---|
| Base model | Same ~743–753B MoE | Same | Identical |
| Post-training | Heavily scaled environments | Earlier stack | 5.3 |
| Terminal-Bench 3.0 | 28.3 | 4.6 | 5.3 (large) |
| DeepSWE v1.1 | 66.9 | 46.2 | 5.3 |
| Internal Code Bench (Max) | 34.5% @ ~75K tokens | 23.4% @ ~96K tokens | 5.3 (perf + efficiency) |
| Agents’ Last Exam | 28.5 | 23.8 | 5.3 |
| AutomationBench | 48.2 | 26.2 | 5.3 |
| CyberGym | 84.5 (SOTA) | 77.2 | 5.3 |
| ExploitBench | 54.4 | 24.4 | 5.3 |
| GDPval-AA v2 | 1769 | 1508 | 5.3 |
| Context / Max output | 1M / 128K | 1M / similar | Same |
| Thinking | Always on (low/high/max) | More flexible previously | 5.3 (always-on) |
| Open weights | ~2 weeks post-launch (planned) | Available (MIT) | 5.2 currently |
| Primary access now | Coding Plan / ZCode | API + weights + Coding Plan | 5.2 more mature |
Introduction: Post-Training Delivers a Generational Jump
In mid-August 2026 the AI community watched another rapid iteration in the open-weights race. Z.ai (the international face of the former Zhipu AI / ChatGLM team) launched GLM-5.3 just 59 days after GLM-5.2. The announcement was unusual in its clarity: the underlying base model remained identical. “Scaling post-training is all we did for GLM-5.3,” the company stated.
That claim is important. For years the dominant narrative has been “bigger models win.” GLM-5.3 demonstrates that aggressive scaling of reinforcement learning environments, long-horizon task diversity, and systems infrastructure can produce outsized gains on hard coding, agentic, and even cybersecurity workloads without a new pre-training run. The numbers are large enough on several difficult benchmarks that independent observers have described the jump as qualitatively different rather than incremental. Overview of GLM-5.3 features, benchmarks, and access notes.
GLM-5.3 vs GLM-5.2: What Actually Changed?
The biggest misconception would be to think GLM-5.3 is simply “GLM-5.2 but bigger.”
It isn't.
The base model remains essentially the same, according to Z.ai. The major innovation is post-training scaling. Z.ai emphasizes three pillars:
- Systems improvements inside slime — Better training
- Environment scaling — Pipelines that synthesize executable, verifiable, long-horizon environments from real professional workflows. Research agents extract task patterns; judge agents verify solvability; verifiers are built without access to reference solutions to reduce reward hacking.
- Continued use of SAO + compaction — Helps gains persist on long trajectories rather than collapsing on short ones. –rollout consistency (log-prob differences controlled to ~1e-7), hierarchical caching, multi-teacher support, workload-aware scheduling, and reported >2.3× end-to-end throughput gains on long-horizon coding RL tasks.
These changes produced measurable improvements across coding, agentic, and cybersecurity suites. Importantly, the largest relative gains appear on the hardest, lowest-baseline tests—exactly the pattern expected when a model is pushed into regimes it previously could not handle reliably.
The result is a model upgrade that is primarily about behavior and capability, rather than simply architecture.
GLM-5.3 vs GLM-5.2: Features Comparison
1. Scaled Post-Training Instead of a New Base Model
Z.ai describes scaled post-training as the entire recipe for GLM-5.3. Research agents turn patterns from real work into runnable tasks with hidden state and multi-step dependencies. A judge agent verifies that each task is solvable. Verifiers are created without seeing the reference solution, then tested against oracle, no-op, and unsolved states to reduce reward shortcuts.
The system still requires human review, and Z.ai explicitly says more autonomous environment generation and verification remain future work. That caveat increases the credibility of the claim: this is a large training pipeline, not a claim that synthetic environments have become fully self-governing.
2. Stronger Long-Horizon Coding
GLM-5.3 improves at repository work, terminal tasks, machine-learning infrastructure, performance optimization, and multi-step software delivery. On Z.ai Code Bench, the private internal evaluation intended to reflect realistic user scenarios, Z.ai reports roughly 50% better coding performance than GLM-5.2.
The efficiency result is as important as the score. At Max effort, GLM-5.3 achieved 34.5% at about 75K output tokens per task, compared with GLM-5.2's 23.4% at 96K. That is an 11.1-point quality gain while producing roughly 22% fewer output tokens. At High effort, GLM-5.3 reached 31.4% using about 50K tokens, ahead of Claude Opus 4.8 at 29.5% with 120K in the same first-party chart. Claude Fable 5 remained higher at 39.5% under Max effort.
3. Emergent Cybersecurity Capability
Z.ai added vulnerability-discovery environments during post-training. According to the launch report, capability grew beyond finding isolated flaws: the model began reasoning across multiple stages of an exploitation chain.
The public scores show both progress and limits. GLM-5.3 leads Z.ai's comparison table on CyberGym at 84.5, versus 77.2 for GLM-5.2. On ExploitBench it more than doubles GLM-5.2, reaching 54.4 versus 24.4, but remains well behind Fable 5 at 78.0 and GPT-5.6 Sol at 76.5. On ExploitGym it completes 105 tasks in a normalized two-hour budget and 130 in six hours, compared with 29 and 39 for GLM-5.2; GPT-5.6 Sol and Fable 5 remain substantially ahead.
These capabilities are dual-use. Organizations should limit model access, sandbox tools, log actions, require authorization for scanning, and keep a human in the loop for vulnerability validation and disclosure.
4. Always-On Reasoning with Three Effort Levels
GLM-5.3 supports low, high, and max reasoning effort. Unlike GLM-5.2, it does not support disabled thinking. Existing integrations that send thinking.type: "disabled" must change the value to enabled before switching the model ID, or Z.ai says the request will fail.
Use low for interactive edits and low-risk transformations, high for substantial repository tasks, and max for difficult coding or security analysis. Higher effort should be evaluated for latency and cost because a stronger answer is not automatically the most economical answer.
5. Better Training Throughput Through slime
GLM-5.3 continues to use slime, Z.ai's open-source framework with Megatron on the training side and SGLang on the rollout side. Z.ai reports additions for top-p masking, top-k and full-vocabulary on-policy distillation, teacher switching, caching, and tighter numerical alignment between training and rollout. The launch report says average log-probability differences were controlled at the 1e-7 level, more than 99.99% lower than earlier setups.
Workload-aware scheduling and load balancing reportedly improved end-to-end reinforcement-learning throughput by more than 2.3 times on long-horizon coding tasks. These are training-infrastructure improvements rather than end-user inference guarantees, but they explain how Z.ai scaled longer and more varied trajectories in a month.
6. Delayed Open Weights for Safety Review
Z.ai calls GLM-5.3 an open-weights model, but the weights were not downloadable on launch day. The company says they will be released two weeks after launch after safety evaluation and hardening. This is a material difference from GLM-5.2, whose MIT-licensed weights are already on Hugging Face.
For most teams, hosted API testing is still the practical first step. The model is large, and open weights do not remove the cost of inference hardware, quantization, serving, monitoring, or security controls.
GLM-5.3 vs GLM-5.2: Benchmark Improvements
The following table uses Z.ai's official launch data. “Change” is a simple calculation from the reported scores. Relative percentages can look dramatic when the GLM-5.2 baseline is low, so the absolute change is shown as well.
| Benchmark | GLM-5.2 | GLM-5.3 | Absolute change | Relative change |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 81.0 | 88.2 | +7.2 | +8.9% |
| Terminal-Bench 3.0 | 4.6 | 28.3 | +23.7 | +515.2% |
| DeepSWE v1.1 | 46.2 | 66.9 | +20.7 | +44.8% |
| NL2Repo | 48.9 | 58.0 | +9.1 | +18.6% |
| ProgramBench Almost Solved | 9.5 | 19.0 | +9.5 | +100.0% |
| FrontierSWE | 67.5 | 78.1 | +10.6 | +15.7% |
| SWE-Marathon v1.1 | 19.4 | 42.5 | +23.1 | +119.1% |
| PostTrainBench | 31.7 | 39.8 | +8.1 | +25.6% |
| CyberGym | 77.2 | 84.5 | +7.3 | +9.5% |
| ExploitBench | 24.4 | 54.4 | +30.0 | +123.0% |
| ExploitGym, 2-hour tasks | 29 | 105 | +76 | +262.1% |
| ExploitGym, 6-hour tasks | 39 | 130 | +91 | +233.3% |
| Toolathlon Verified | 59.9 | 73.0 | +13.1 | +21.9% |
| AutomationBench v1.0.6 | 26.2 | 48.2 | +22.0 | +84.0% |
| Agents' Last Exam CLI | 23.8 | 28.5 | +4.7 | +19.7% |
| HLE with tools | 54.7 | 62.5 | +7.8 | +14.3% |
| GDPval-AA v2, Elo | 1508 | 1769 | +261 | +17.3% |
Benchmark Improvements: GLM-5.3 vs GLM-5.2 and Competitors
All numbers below are vendor-reported from the official Z.ai blog (with methodology notes on harnesses, context lengths, and sampling). Independent verification will follow once weights and broader access are available.
Coding Benchmarks
| Benchmark | GLM-5.3 | GLM-5.2 | Notes / Competitors |
|---|---|---|---|
| Terminal Bench 2.1 | 88.2 | 81.0 | Competitive with top closed models |
| Terminal Bench 3.0 | 28.3 | 4.6 | Open-source SOTA; vs Fable 5 ~33.7, GPT-5.6 Sol ~34.6 |
| DeepSWE v1.1 | 66.9 | 46.2 | Strong jump |
| NL2Repo | 58.0 | 48.9 | — |
| ProgramBench Almost Solved | 19.0 | 9.5 | — |
| FrontierSWE | 78.1 | 67.5 | — |
| SWE-Marathon v1.1 | 42.5 | 19.4 | — |
| Z.ai Code Bench (internal, Max effort) | ~34.5% completion @ ~75K tokens | ~23.4% @ ~96K tokens | ~50% overall improvement in coding feel; higher efficiency |
At High effort, GLM-5.3 reached 31.4% completion with ~50K tokens, surpassing Claude Opus 4.8 (29.5% with 120K tokens) on the internal bench, while still trailing Claude Fable 5 (39.5% at Max).
Cybersecurity Benchmarks
| Benchmark | GLM-5.3 | GLM-5.2 | Competitors (approx.) |
|---|---|---|---|
| CyberGym | 84.5% | 77.2% | Mythos 5: 83.8%, GPT-5.6 Sol: 83.6% (SOTA) |
| ExploitBench | 54.4% | 24.4% | More than doubles prior; closed models higher (Mythos 5 ~78%, GPT-5.6 Sol ~76.5%) |
| ExploitGym (2h/6h) | 105 / 130 | 29 / 39 | Mythos 5 still ahead (181/247) |
Gains are largest further up the exploitation chain. In real-world testing with Chinese security teams, the model (building on GLM-5.2 work) identified 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues. Some flaws dated back ~40 years (oldest ~1981). Findings are tracked in the public Z.ai Security Disclosure Ledger .
Agentic & Other Benchmarks
| Benchmark | GLM-5.3 | GLM-5.2 | Notes |
|---|---|---|---|
| Toolathlon Verified | 73.0 | 59.9 | — |
| AutomationBench v1.0.6 | 48.2 | 26.2 | Large gain |
| Agents’ Last Exam (ALE-CLI) | 28.5 | 23.8 | Open-source competitive |
| HLE w/ Tools | 62.5 | 54.7 | — |
| GDPval-AA v2 | 1769 | 1508 | Covers 44 professions |
These results demonstrate clear progress on long-horizon, multi-step professional tasks.
Token Efficiency, Thinking Modes, and Practical Behavior
A quiet but important improvement is token efficiency. On the internal Code Bench, higher completion rates arrive with fewer output tokens. This matters for cost and latency in production agent loops.
GLM-5.3 always enables thinking. Three effort levels are supported: low, high, and max (default and recommended for coding is max). Disabling thinking is no longer supported; applications that previously set thinking.type: "disabled" must migrate.
Context remains a solid 1M tokens with maximum output up to 128K. The architecture is unchanged from GLM-5.2, so hardware sizing for future self-hosting can be estimated from existing GLM-5.2 experience (quantized footprints still large given the total parameter count).
For developers who want to experiment immediately or maintain flexibility across models, unified gateways are useful. CometAPI lists GLM-series models (including GLM-5.3 under the model ID glm-5.3 as it becomes available) with OpenAI-compatible endpoints, competitive rates , usage-based billing, and the ability to switch between GLM-5.2, GLM-5.3, and dozens of other frontier and open models without rewriting integration code. This is particularly practical for A/B testing coding agents, measuring real cost-per-successful-task, and routing traffic based on workload.
Who Should Move to GLM-5.3 and How to Evaluate
Teams building coding agents, long-horizon automation, repository-scale engineering workflows, or defensive security tooling should prioritize evaluation. The combination of higher completion rates, better token efficiency on complex tasks, and stronger multi-step agency is material.
Recommended evaluation path:
- Run the same internal tasks (or a fixed harness) on GLM-5.2 and GLM-5.3 (or via Coding Plan) at matched effort levels.
- Measure not only pass rate but tokens, wall-clock time, and human intervention rate.
- Use a unified API layer such as CometAPI to keep the comparison frictionless and to retain the option to fall back or route selectively.
- For security-sensitive workloads, wait for the full safety-hardened open weights and review disclosure practices.
Self-hosting will become attractive once weights ship, especially for organizations that already run GLM-5.2 infrastructure.
Conclusion: Post-Training as the New Scaling Frontier
GLM-5.3 is one of the clearest recent demonstrations that post-training scale—more realistic environments, better RL infrastructure, and sustained compute on long trajectories—can produce capability jumps that previously seemed to require new base models. The coding and agentic gains are large; the cyber results were largely emergent and already competitive or leading on discovery-oriented benchmarks.
For practitioners the practical takeaway is straightforward: test the model on your actual workloads, measure cost per successful outcome rather than raw leaderboard position, and keep integration flexible. Platforms such as CometAPI simplify that process by offering a single key, OpenAI-compatible interface, and easy switching across the GLM series and competing models while you decide how aggressively to adopt the new capabilities.