GLM-5.3 and GLM-5 are two open-weight large language models from Z.ai (Zhipu), released six months apart in 2026, and code written for GLM-5 does not always run unchanged on GLM-5.3. GLM-5 arrived on 12 February 2026 as the first model of the GLM-5 generation; GLM-5.3 followed on 18 August 2026 after two intermediate releases, GLM-5.1 and GLM-5.2. This page puts the two side by side: API identifiers, release dates, parameter counts, context window, output limits, prices, reasoning controls, licences, availability, provider-reported benchmarks and lifecycle status. It is written for developers planning a migration, buyers estimating cost and architects deciding which GLM identifier to pin. Models are listed newest first, an objective order that implies no ranking, and every fact comes from Z.ai's own documentation or Hugging Face model cards.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Z.ai's public documentation as of November 2026. Model details change, so confirm against Z.ai's current documentation.
Key takeaways
- Context grew from 200K to 1M tokens. Z.ai's GLM-5 page lists a 200K context length; the GLM-5.3 page lists 1M tokens. Both list 128K maximum output tokens.
- GLM-5.3 costs 40% more per input token. The pricing page lists GLM-5 at $1 input, $0.2 cached input and $3.2 output per million tokens, and GLM-5.3 at $1.4, $0.26 and $4.4.
- Reasoning can no longer be switched off. GLM-5 accepts thinking.type enabled or disabled; GLM-5.3 accepts only enabled and adds reasoning_effort with low, high and max (default max).
- The licence changed. The GLM-5 model card states the MIT licence; GLM-5.3 uses a licence named glm-5.3, which adds a security review condition for very large Model as a Service operators.
- Both remain on the price list. Z.ai's pricing page still lists GLM-5 alongside GLM-5.3, and neither the release notes nor the model pages show a deprecation date for either model.
What is the GLM-5 line?
GLM is Zhipu's model family, offered internationally under the Z.ai brand through the Z.ai API Platform and as open weights from the zai-org organisation on Hugging Face. The GLM-5 line has four releases so far: GLM-5 (February 2026), GLM-5.1 (April 2026), GLM-5.2 (June 2026) and GLM-5.3 (August 2026). This page compares the first and the latest; the two intermediate versions appear in the timeline below where they explain a change.
Both models use a Mixture-of-Experts design, carry the glm_moe_dsa architecture tag on Hugging Face and cite the same technical report, "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763). Z.ai's GLM-5.3 page states that GLM-5.3 uses the same base model as GLM-5.2, with its improvements coming from post-training. For the full line-up, see the Zhipu Z.ai GLM Models hub, and for the GLM-5.3 detail, including the separate multimodal GLM-5.3-Flash, see GLM-5.3 and GLM-5.3 Flash. Background on sparse expert models is in Mixture of Experts Explained.
Both models at a glance
As of 2026-11-15. Sources: Z.ai model pages, pricing page, release notes and Hugging Face model cards.
| Field | GLM-5.3 | GLM-5 |
|---|---|---|
| API ID | glm-5.3 | glm-5 |
| Released | 2026-08-18 | 2026-02-12 |
| Status on Z.ai pages | Described as Z.ai's latest flagship model | Listed and priced; positioned as a Foundation Model |
| Context window | 1M tokens | 200K tokens |
| Max output | 128K tokens | 128K tokens |
| Input / output modalities | Text / text | Text / text |
| Knowledge cutoff | Not stated by Z.ai on the pages checked (2026-11-15) | Not stated by Z.ai on the pages checked (2026-11-15) |
| Parameters | 753B total (card metadata); active not stated by Z.ai | 744B total, 40B active (model page and card text); card metadata lists 754B |
| Tensor types on Hugging Face | BF16, F8_E4M3, F32 | BF16, F32 |
| Open weights | Yes: zai-org/GLM-5.3 | Yes: zai-org/GLM-5 |
| Licence | glm-5.3 (publisher licence) | MIT |
| Input price per 1M tokens | $1.4 | $1 |
| Output price per 1M tokens | $4.4 | $3.2 |
The GLM-5 card does not explain the gap between the 744B figure in its introduction and the 754B in its file metadata; both are shown so readers can match either figure they meet elsewhere.
When was each model released?
Dates from Z.ai's release notes, as of 2026-11-15, oldest first.
| Date | Release | What the release note says |
|---|---|---|
| 2026-02-12 | GLM-5 | Designed for complex system engineering and long-range agent tasks; integrates DeepSeek Sparse Attention for higher token efficiency |
| 2026-04-07 | GLM-5.1 | Built for long-horizon tasks, with independent runs of up to 8 hours |
| 2026-06-16 | GLM-5.2 | Adds a 1M token context window |
| 2026-08-18 | GLM-5.3 | A stated 50% gain over GLM-5.2 on Z.ai Code Bench, plus vulnerability-finding results |
The 1M context window therefore arrived with GLM-5.2, not GLM-5.3, and the move from GLM-5.2 to GLM-5.3 was a post-training change on the same base model. Anyone jumping straight from GLM-5 to GLM-5.3 inherits both steps at once. For how point releases are named across providers, see LLM Naming and Versioning Explained.
How much does each model cost?
As of 2026-11-15. USD per million tokens. Source: Z.ai pricing page.
| Item | GLM-5.3 | GLM-5 | Notes |
|---|---|---|---|
| Input | $1.4 | $1 | Standard API price |
| Cached input | $0.26 | $0.2 | Charged when a cached prefix is reused |
| Cached input storage | Limited-time free | Limited-time free | No end date shown on the pricing page |
| Output | $4.4 | $3.2 | Reasoning tokens are output tokens |
| Batch pricing | Not stated by Z.ai on the pages checked (2026-11-15) | Not stated by Z.ai on the pages checked (2026-11-15) | No batch tier on the pricing page |
GLM-5.1 and GLM-5.2 carry the same prices as GLM-5.3, so the price step happened at GLM-5.1, not at the latest release.
Worked example: a request with 100,000 input tokens and 10,000 output tokens costs $0.10 plus $0.032, about $0.13, on GLM-5, and $0.14 plus $0.044, about $0.18, on GLM-5.3. The real difference can be larger or smaller because GLM-5.3 reasons on every call at max effort unless told otherwise, which changes the output token count. Z.ai's own Code Bench figure illustrates this: at max effort GLM-5.3 used about 75K output tokens per task against 96K for GLM-5.2. Measure output tokens on your own prompts before forecasting. See Model Pricing Explained and the LLM Pricing Tracker.
What changed between models?
| Area | GLM-5 | GLM-5.3 | Effect on an upgrade |
|---|---|---|---|
| Reasoning switch | thinking.type accepts disabled or enabled (default enabled) | thinking.type accepts enabled only | Requests sending disabled fail on GLM-5.3 |
| Reasoning depth | Multiple thinking modes mentioned, not named | reasoning_effort: low, high, max (default max) | Default calls reason at the deepest level |
| Context | 200K | 1M | Longer inputs accepted; cost per call can rise with them |
| Weights precision | BF16 and F32 tensors | BF16, F8_E4M3 and F32 tensors | An FP8 tensor type is listed on the card |
| Licence | MIT | glm-5.3 licence | Legal review needed before redistribution or hosting |
| API endpoints listed | OpenAI-compatible base URL api.z.ai/api/paas/v4/ | OpenAI Chat Completion, OpenAI Responses and Anthropic Messages compatible base URLs | More client protocols documented |
Z.ai's migration note, written for GLM-5.2 users but equally relevant to GLM-5 code, says applications that send thinking.type disabled must change it to enabled and set reasoning_effort to low before switching the model ID to glm-5.3, otherwise requests fail. The same page notes that users with a previous GLM Coding Plan subscription, including expired ones, can currently reach the model API only through the OpenAI Chat Completion compatible protocol. For background on reasoning controls, see Reasoning Models Explained.
What workloads does Z.ai position each model for?
- GLM-5.3: the model page describes it as "Z.ai's latest flagship model", focused on complex software engineering and agent capabilities, and recommends max effort for complex tasks such as coding.
- GLM-5: the release note says it is designed for complex system engineering and long-range agent tasks; its model page labels it a Foundation Model.
Both positions centre on coding and agent work. The difference is in the evaluations Z.ai chose to publish: the GLM-5 card reports broad suites such as GPQA-Diamond, AIME 2026 I and BrowseComp, while the GLM-5.3 card concentrates on coding, long-running engineering and cyber benchmarks, with a section headed "Emergent Cyber Capability".
What is the context window of each model?
| Limit | GLM-5.3 | GLM-5 | Source |
|---|---|---|---|
| Context window | 1M tokens | 200K tokens | Z.ai model pages |
| Maximum output | 128K tokens | 128K tokens | Z.ai model pages |
| Reasoning control | reasoning_effort low, high, max | thinking.type enabled or disabled | Z.ai model pages |
| Rate limits | Not stated by Z.ai on the pages checked (2026-11-15) | Not stated by Z.ai on the pages checked (2026-11-15) | Z.ai model pages |
The Hugging Face cards also quote context lengths in benchmark notes; those are evaluation settings, not product limits. Hosts that serve the open weights set their own limits. See Context Windows Explained.
Capabilities and tools compared
| Capability | GLM-5.3 | GLM-5 |
|---|---|---|
| Thinking mode | Listed, always on | Listed, can be disabled |
| Streaming output | Listed | Listed |
| Function calling | Listed | Listed |
| Context caching | Listed | Listed |
| Structured output (JSON) | Listed | Listed |
| Image or video input | Text-only input | Text-only input |
| SDKs named | zai-sdk and the OpenAI Python SDK | zai-sdk for Python (0.2.3) and Java (0.3.5) |
The five listed capabilities are the same on both pages; the change is how reasoning is controlled. Neither model accepts images: Z.ai's first natively multimodal GLM-5 model is GLM-5.3-Flash, a different, smaller model covered on the GLM-5.3 profile.
Where is each model available?
| Channel | GLM-5.3 | GLM-5 |
|---|---|---|
| Z.ai API Platform | glm-5.3 | glm-5 |
| Open weights | zai-org/GLM-5.3 on Hugging Face | zai-org/GLM-5 on Hugging Face |
| Self-hosting engines on the card | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, Ascend NPU options (vLLM-Ascend, xLLM, SGLang), Docker Model Runner | vLLM 0.19.0+, SGLang 0.5.10+, KTransformers 0.5.3+, Transformers, xLLM 0.8.0+ |
| Third-party clouds | Not stated by Z.ai on the pages checked (2026-11-15) | Not stated by Z.ai on the pages checked (2026-11-15) |
Running the weights yourself moves hosting, patching and access control to your team. See the Self-Hosted LLM Security Guide and the Open-Weight Model Security Guide.
Is GLM-5.3 or GLM-5 open source?
Both publish weights. The GLM-5 card states the MIT licence. The GLM-5.3 card names a licence called glm-5.3, whose text grants free rights to use, copy, modify, distribute, sell, fine-tune and build derivative works, with three conditions:
- Large MaaS operators: if the licensee and its affiliates run a Model as a Service business with combined revenue above 10 billion US dollars in any consecutive 12-month period, the licensee must pass Z.ai's security review before any commercial use.
- MaaS definition: giving a third party inference or fine-tuning access with meaningful control over inputs, parameters or training data; it excludes end-user products that embed the model in specific features and the relaying of requests to models hosted by others.
- Notices and law: the copyright and permission notice must be kept, and use must comply with applicable laws.
For most organisations the practical change is a licence review step. For how open-weight licences differ, see Open-Weight vs Closed Models Explained and the open-weight model licence tracker.
Model cards and system cards
| Document | GLM-5.3 | GLM-5 |
|---|---|---|
| Model card | huggingface.co/zai-org/GLM-5.3 | huggingface.co/zai-org/GLM-5 |
| Documentation page | docs.z.ai/guides/llm/glm-5.3 | docs.z.ai/guides/llm/glm-5 |
| Technical report | GLM-5 report, arXiv 2602.15763 (cited) | GLM-5 report, arXiv 2602.15763 |
| System card | Not stated by Z.ai on the pages checked (2026-11-15) | Not stated by Z.ai on the pages checked (2026-11-15) |
| Dedicated safety section | None on the card; evaluation safeguards described | None on the card |
The GLM-5 page states training data rose from 23T to 28.5T tokens against the previous generation and that DeepSeek Sparse Attention was integrated for the first time. The GLM-5.3 card describes anti-cheating measures for its evaluations: domain allow-lists, removal of Git information and rule-based and LLM-based checks for behaviours such as unauthorised pip or curl operations.
Benchmarks
As of 2026-11-15. Provider-reported figures only, taken from the two Hugging Face model cards (GLM-5 card published with the February 2026 release; GLM-5.3 card with the August 2026 release). Rows appear only where both cards report the same evaluation. Settings and harness versions can differ between cards, so treat the gap as indicative.
| Evaluation | GLM-5.3 | GLM-5 | Source and date |
|---|---|---|---|
| Humanity's Last Exam, with tools | 62.5 | 50.4 | Z.ai model cards, August and February 2026 |
| CyberGym | 84.5 | 43.2 | Z.ai model cards, August and February 2026 |
Other evaluations appear on only one card. The GLM-5 card reports SWE-bench Verified 77.8 and Terminal-Bench 2.0 56.2; the GLM-5.3 card reports Terminal Bench 2.1 (88.2) and 3.0 (28.3) and DeepSWE v1.1 (66.9), with GLM-5.2 as its prior-version comparator rather than GLM-5. The two Terminal Bench versions are different tests, so their numbers are not comparable. Independent leaderboard results are not included because none were checked against the leaderboards' own pages for this comparison. See How to Read LLM Benchmarks.
The cyber gain is notable for security teams: the GLM-5.3 release note says the model identified 2,436 vulnerabilities, including 1,097 medium and high severity, in real-world work, and its card reports large rises on exploitation benchmarks over GLM-5.2. Open weights mean that capability is available to anyone who downloads them.
Deprecations and retirements
| Model or item | Details | Source |
|---|---|---|
| GLM-5.3 | Current; no deprecation date stated by Z.ai on the pages checked (2026-11-15) | Z.ai model page; release notes |
| GLM-5 | Still listed on the pricing page; no deprecation notice on its model page or in the release notes checked (2026-11-15) | Z.ai pricing page; GLM-5 model page |
| Open weights | Downloaded checkpoints remain usable whatever the API lifecycle | Hugging Face model cards |
Z.ai publishes no general deprecation schedule on the pages checked, so the absence of a date is not a guarantee of availability. Pin the identifier and watch the release notes.
Data handling and security
The same terms apply to both models on the Z.ai API. Z.ai's privacy policy names Jingsheng Hengxing Technology Pte. Ltd. in Singapore as controller for individual users and says the services are generally provided from Singapore. For API customers, the data processing agreement summarised in the policy makes the customer the controller and Z.ai the processor, and states that content customers input or generate is processed in real time and is not saved on Z.ai's servers; processing is limited to providing the API service and following the customer's written instructions. Confirm these terms in your own agreement.
The largest operational risk is the key, not the model. Z.ai's Anthropic Messages compatible endpoint means coding agents can be pointed at GLM-5.3 by changing a base URL and API key, so treat that key as a credential with spend authority. See the AI Coding Agents Security Guide, the LLMjacking Guide and How to set spend limits on LLM API keys.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Inputs above 200K tokens | GLM-5.3 only (1M); GLM-5 lists 200K |
| List price per token | GLM-5: $1 / $3.2; GLM-5.3: $1.4 / $4.4 |
| Ability to turn reasoning off | GLM-5 only |
| Licence | GLM-5: MIT; GLM-5.3: glm-5.3 licence with a MaaS review condition |
| FP8 tensors on the card | GLM-5.3 lists F8_E4M3; GLM-5 lists BF16 and F32 |
| API protocols documented | GLM-5.3 lists OpenAI Chat Completion, OpenAI Responses and Anthropic Messages compatible URLs |
Run your own evaluation on representative prompts and count output tokens before switching. For other Chinese model families side by side, see DeepSeek vs Qwen vs Kimi vs GLM, and for the wider field, Frontier AI Models Compared.
Practitioner checklist
- Search code and agent configs for thinking.type disabled and change it to enabled with reasoning_effort low before moving to glm-5.3.
- Pin the model identifier (glm-5 or glm-5.3) explicitly and record it in the AI inventory.
- Re-measure output tokens on real prompts after the switch, since reasoning runs at max by default.
- Have legal review the glm-5.3 licence before hosting, redistributing or fine-tuning the weights.
- Recalculate budgets at $1.4 / $4.4 per million tokens, and do not plan around the free cached storage offer, which has no stated end date.
- Store Z.ai API keys in a secrets manager and set spend limits, especially where coding agents use the Anthropic-compatible endpoint.
- Watch Z.ai's release notes for lifecycle changes, since no deprecation schedule is published.
Standards and references
- Z.ai: GLM-5.3 model guide (accessed 2026-11-15)
- Z.ai: GLM-5 model guide (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
- Z.ai: Release notes (accessed 2026-11-15)
- Z.ai: GLM-5.3 model card on Hugging Face (accessed 2026-11-15)
- Z.ai: GLM-5 model card on Hugging Face (accessed 2026-11-15)
- Z.ai: GLM-5.3 licence (accessed 2026-11-15)
- Z.ai: Privacy policy (accessed 2026-11-15)
Related NHI Mgmt Group resources: Zhipu Z.ai GLM Models, GLM-5.3 and GLM-5.3 Flash, DeepSeek vs Qwen vs Kimi vs GLM, Open-Weight Frontier Models Compared, LLM Pricing Tracker.