GLM-5.3 is Z.ai’s August 14, 2026 answer to a simple question: what happens if you keep scaling post-training on the same base? It uses the same 753B-parameter base model as GLM-5.2 and changes nothing else — more training environments, more diverse long-horizon tasks, more RL compute. The result is a jump of the size usually reserved for a new generation: Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9.
It keeps the 1M-token context and 128K output limit, ships as open weights, and does not change the price. The other headline is cyber: Z.ai says GLM-5.3 posted the best CyberGym score at launch and more than doubled GLM-5.2 on exploitation benchmarks.
Benchmarks
Vendor-reported by Z.ai at launch (August 14, 2026), at maximum effort in Claude Code-style harnesses. The comparison columns are as published by Z.ai (its Qwen3.8-Max column is omitted for width).
| Benchmark | GLM-5.3 | GLM-5.2 | Kimi K3 | DeepSeek V4-Pro-0813 | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.2 | 81.0 | 88.3 | 87.9 | 85.0 | 88.0 | 88.8 |
| Terminal-Bench 3.0 | 28.3 | 4.6 | 17.4 | — | 21.1 | 33.7 | 34.6 |
| DeepSWE v1.1 | 66.9 | 46.2 | 67.5 | 62.7 | 58.0 | 69.7 | 72.7 |
| ProgramBench | 19.0 | 9.5 | 17.5 | — | 15.5 | 33.0 | 23.0 |
| CyberGym | 84.5 | 77.2 | 80.0 | 83.3 | 78.1 | 83.8 | 83.6 |
| ExploitBench | 54.4 | 24.4 | 32.2 | — | 40.0 | 78.0 | 76.5 |
| Toolathlon Verified | 73.0 | 59.9 | 76.5 | 74.1 | 76.2 | 74.7 | 74.9 |
| AutomationBench | 48.2 | 26.2 | 46.7 | 43.2 | 41.0 | 46.2 | 45.8 |
| Agents’ Last Exam (CLI) | 28.5 | 23.8 | 27.6 | 25.7 | 25.7 | 23.8 | 28.6 |
| HLE with tools | 62.5 | 54.7 | 59.8 | 60.0 | 57.9 | 63.9 | 64.5 |
The Fable 5 column is “Fable 5 (w/ fallback)” in Z.ai’s table; in its text Z.ai attributes the cyber figures in that column (CyberGym 83.8, ExploitBench 78.0) to Mythos 5. Terminal-Bench 3.0 uses an avg@3 protocol in Claude Code with 600-turn and 10-hour caps.
On its private Z.ai Code Bench, GLM-5.3 reaches 34.5% at max effort using about 75K output tokens per task (GLM-5.2: 23.4% at 96K). At high effort it hits 31.4% at around 50K tokens, ahead of Claude Opus 4.8’s 29.5% at 120K — but behind Claude Fable 5 at 39.5%.
Cyber capability
Z.ai says the cyber results were a surprise: vulnerability-discovery data added to post-training kept paying off as training scaled, and the model began to plan full exploitation chains. On CyberGym it scores 84.5%, and on ExploitGym it completes 105 tasks in two hours and 130 in six (GLM-5.2: 29 and 39), though Mythos 5 completes 181 and 247.
Z.ai also reports running its models against real codebases with security teams in China: 2,436 vulnerabilities across 269 projects, 1,097 of them medium or higher, tracked in a public Security Disclosure Ledger (53 disclosed and 2,383 under embargo at launch). Treat these as vendor claims — the ledger is the place to check them.
Pricing
| Per 1M tokens | Price |
|---|---|
| Input | $1.40 |
| Cached input | $0.26 |
| Output | $4.40 |
Same as GLM-5.2 and GLM-5.1. GLM-5.3-Flash and FlashX are far cheaper — see GLM-5.3-Flash.
Specs
| Spec | Value |
|---|---|
| API model | glm-5.3 |
| Base | Same as GLM-5.2 — 753B-parameter MoE (glm_moe_dsa) |
| Context window | 1M tokens; max output 128K |
| Modalities | Text in, text out |
| Reasoning | Always on; reasoning_effort low / high / max (default max) |
| Weights | Hugging Face zai-org/GLM-5.3 (BF16 / FP8) |
| License | GLM-5.3 License — MIT-style, with a security-review condition for Model-as-a-Service operators above $10B revenue |
| Serving | SGLang, vLLM, Transformers, KTransformers, Unsloth, Ascend NPU stacks |
| Protocols | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages |
How it compares
- vs GLM-5.2: same base, same price, far better agentic scores — and mandatory thinking.
- vs Kimi K3: within a point on DeepSWE v1.1 and Terminal-Bench 2.1, ahead on Terminal-Bench 3.0 and CyberGym, behind on Toolathlon.
- vs DeepSeek V4.1-Flash: DeepSeek’s own table has V4.1-Flash ahead on DeepSWE v1.1 (74.2 vs 66.9) while GLM-5.3 leads on Terminal-Bench 4.0 (37.9 vs 31.2); the harnesses differ.
- vs closed frontier: it closes in on GPT-5.6 Sol and Fable 5 on several rows but trails the newer GPT-6 and Claude Opus 5.5 generation.
Limitations
- Text only at the flagship tier.
- Vendor-reported numbers, run in Z.ai’s own Claude Code-based harnesses.
- Mandatory thinking breaks clients that disable it.
- Cyber gap remains: Z.ai itself notes the gap to the closed frontier widens further up the exploitation chain.
For the wider picture, see the 2026 LLM guide.
Best for
- Long-horizon autonomous coding and engineering agents
- Vulnerability discovery and security research in sandboxed environments
- Repo-scale refactors and migrations at a 1M context
- Self-hosted open-weight coding stacks in Claude Code, OpenCode and ZCode
Pros & cons
Strengths
- Z.ai calls it the most capable open-weight coding model, with open-source-best results on Terminal-Bench 3.0 (28.3) and Agents' Last Exam (28.5)
- A big jump over GLM-5.2 from post-training alone — DeepSWE v1.1 46.2 to 66.9, AutomationBench 26.2 to 48.2 — on the same base model and the same price
- Weights are on Hugging Face under an MIT-style license with one narrow carve-out (Model-as-a-Service providers above $10B revenue need a security review)
- Token-efficient at high effort — Z.ai's in-house bench shows 31.4% at about 50K output tokens, ahead of Claude Opus 4.8 at 29.5% with 120K
Limitations
- Text-only at the flagship tier; the multimodal member of the family is the smaller GLM-5.3-Flash
- Behind closed frontier models on the hardest rows — Terminal-Bench 3.0 28.3 vs 34.6 for GPT-5.6 Sol, and ExploitBench 54.4 vs 78.0 for Mythos 5
- Thinking is mandatory — disabling it now fails, so old requests must be migrated before switching the model ID
- A 753B-parameter MoE still needs a multi-GPU node to self-host
Frequently asked questions
When was GLM-5.3 released?
Z.ai launched GLM-5.3 on August 14, 2026 on its API and the GLM Coding Plan. It said the weights would follow two weeks later after safety evaluation; they are now on Hugging Face as `zai-org/GLM-5.3`.
Is GLM-5.3 a new model or an update to GLM-5.2?
It uses the same base model as GLM-5.2. Z.ai says every gain comes from post-training — more environments, more diverse tasks and more RL compute on the same stack.
How much does GLM-5.3 cost?
$1.40 per 1M input tokens, $0.26 per 1M cached input tokens and $4.40 per 1M output tokens — unchanged from GLM-5.2. Cached input storage is free for a limited time.
What license are the GLM-5.3 weights under?
A custom GLM-5.3 License that is MIT-style. The one condition beyond attribution is that a Model-as-a-Service operator with more than $10 billion in 12-month revenue must pass a Z.ai security review before commercial use.
Can I turn thinking off?
No. GLM-5.3 always reasons. It supports `reasoning_effort` of low, high or max (default max). If your app sends `thinking.type: disabled`, change it to `enabled` and set low effort before switching the model ID to `glm-5.3`, or the request fails.