What Is GLM-5.3 in 2026?
GLM-5.3 is z.ai's flagship text model in 2026, built on the same base model as GLM-5.2 with every improvement coming from post-training. z.ai positions it for complex coding, long-horizon agent tasks and cybersecurity work. It is text-only, has a 1 million token context window and a 128K output limit, and always reasons, at low, high or max effort.
This post first covered GLM-5.3 launching without downloadable weights. z.ai had promised them about two weeks after launch, and the GLM-5.3 and GLM-5.3-Flash repositories appeared on Hugging Face on 25 August under a custom GLM-5.3 License. The sections below reflect that.
| Attribute | What z.ai has published |
|---|---|
| Launch | Blog post dated 14 August; docs release notes list 18 August |
| Base model | Same as GLM-5.2; gains from post-training only |
| Parameters | About 753 billion, per the Hugging Face model files |
| Context and output | 1M-token context, 128K maximum output, text input only |
| Reasoning | Always on; effort levels low, high and max (default max) |
| Open weights | On Hugging Face since 25 August, including BF16 versions |
| Licence | Custom GLM-5.3 License with a security review condition for very large MaaS providers |
| API price | $1.40 input, $4.40 output per 1M tokens |
| Coding Plan | Included in Lite, Pro and Max |
When Were GLM-5.3 and Its Weights Released?
z.ai's blog announced GLM-5.3 in a post dated 14 August 2026, and its documentation release notes list the model under 18 August. At launch z.ai said the weights would follow in two weeks, once safety evaluation and hardening were done. The Hugging Face repositories were created on 25 August, and GLM-5.3-Flash followed on 26 August.
| Date (2026) | Event | Source |
|---|---|---|
| 14 August | GLM-5.3 launch blog post | z.ai blog |
| 18 August | GLM-5.3 listed in release notes | z.ai docs |
| 25 August | GLM-5.3 and GLM-5.3-Flash weights published | Hugging Face (zai-org) |
| 26 August | GLM-5.3-Flash listed in release notes | z.ai docs |
The staged release was explained by cybersecurity: z.ai says the model's cyber capability grew faster than expected as post-training scaled, so it held the weights back for extra safety work.
What Does the GLM-5.3 Licence Allow?
GLM-5.3's weights use a custom GLM-5.3 License rather than a standard open-source licence. It grants free rights to use, copy, modify, distribute, sell, deploy and fine-tune the model, provided the copyright notice is kept. The extra condition applies only to Model as a Service businesses with group revenue above $10 billion over 12 months, which must pass Z.AI's security review before commercial use.
- Most businesses: can self-host, fine-tune and use GLM-5.3 commercially under the licence terms.
- Embedding in your own product: the licence excludes end-user products where model capability is embedded in specific features from its Model as a Service definition.
- Very large API providers: must pass Z.AI's security review first.
- Always read the licence yourself or with counsel before deploying for clients.
Our GLM-5.2 self-hosting guide covers hardware planning; z.ai's model card lists SGLang, vLLM, TokenSpeed, Transformers and KTransformers as supported serving options for GLM-5.3.
What Do the GLM-5.3 Benchmarks Show in 2026?
z.ai's own benchmark table shows large gains over GLM-5.2 on coding and agent evaluations, and a lead on the CyberGym vulnerability discovery benchmark. Every figure is vendor-reported, so treat it as a claim to test. z.ai also reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench.
| Benchmark (z.ai-reported) | GLM-5.3 | GLM-5.2 | Best listed comparison |
|---|---|---|---|
| Terminal Bench 3.0 | 28.3 | 4.6 | GPT-5.6 Sol 34.6 |
| DeepSWE v1.1 | 66.9 | 46.2 | GPT-5.6 Sol 72.7 |
| AutomationBench v1.0.6 | 48.2 | 26.2 | GLM-5.3 highest listed |
| CyberGym | 84.5 | 77.2 | GLM-5.3 highest listed |
| ExploitBench | 54.4 | 24.4 | Fable 5 78.0 |
| GDPval-AA v2 | 1769 | 1508 | GLM-5.3 highest listed |
Two readings matter. The Terminal Bench 3.0 jump is large, yet the score still means most tasks fail. And on security, GLM-5.3 leads at finding vulnerabilities but trails Fable 5 and GPT-5.6 Sol on exploitation benchmarks in z.ai's own table.
How Much Does GLM-5.3 Cost in 2026?
GLM-5.3's API price on z.ai's pricing page is $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens, the same list price as GLM-5.2. GLM-5.3-Flash costs $0.15 input and $0.50 output per million tokens. Both are included in every GLM Coding Plan tier, which starts at $18 a month.
| Option | Price |
|---|---|
| GLM-5.3 API | $1.40 input, $0.26 cached input, $4.40 output per 1M tokens |
| GLM-5.3-Flash API | $0.15 input, $0.03 cached input, $0.50 output per 1M tokens |
| GLM Coding Plan | Lite $18, Pro $80, Max $168 a month (monthly billing) |
| Self-hosting | Your own hardware costs under the GLM-5.3 License |
See our GLM Coding Plan pricing guide for plan limits and discounts.
What Should Teams Running GLM-5.2 Do in 2026?
Most teams should test GLM-5.3 now. The API price is unchanged, the weights are available, and z.ai reports clear coding gains. Two practical points need attention: GLM-5.3 always reasons, so requests that disabled thinking must be updated, and Coding Plan requests for GLM-5.2 are routed to GLM-5.3 automatically, which affects version pinning.
| Your situation | Recommended step |
|---|---|
| Coding Plan subscriber | You already get GLM-5.3; re-test key prompts since GLM-5.2 requests are routed to it |
| API integrator | Switch the model ID to glm-5.3 after updating thinking settings; compare cost per task |
| Self-hosting GLM-5.2 | Review the GLM-5.3 License, then test the new weights on your hardware |
| Need a fixed version | Pin GLM-5.2 via the standard API or your own deployment, not the Coding Plan |
z.ai's docs say GLM-5.3 does not support disabling reasoning. If your code sends reasoning switched off, switch it on with low effort before switching the model ID, or requests will fail.
Key Takeaways for 2026
- GLM-5.3 launched in mid-August 2026 (z.ai blog 14 August; release notes 18 August).
- The open weights followed on Hugging Face on 25 August under a custom GLM-5.3 License.
- API pricing matches GLM-5.2 at $1.40 input and $4.40 output per million tokens.
- Benchmarks are vendor-reported; GLM-5.3 leads on CyberGym and AutomationBench in z.ai's table.
- Update thinking settings before migrating, and note Coding Plan routes GLM-5.2 requests to GLM-5.3.