GLM-5.3 and GLM-5.3 Flash: Specs, Pricing, Context Window and Availability Facts

GLM-5.3 is an open-weight large language model from Z.ai (Zhipu), released on 18 August 2026, which Z.ai's documentation describes as its latest flagship model, focused on complex software engineering and agent capabilities; GLM-5.3-Flash, its smaller multimodal sibling, followed on 26 August 2026. This profile collects the published facts about both: API identifiers, parameter counts, context window, output limits, prices, reasoning controls, modalities, availability, licences, provider-reported benchmarks and lifecycle. It also covers the GLM-5.3-FlashX and GLM-5.3-FlashX-Legal variants that Z.ai documents on the Flash page. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Where Z.ai publishes nothing on a point, the page says so.

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 and model cards as of November 2026. Model details change, so confirm against Z.ai's current documentation.

Key takeaways

• Released 18 and 26 August 2026. Z.ai's release notes record GLM-5.3 on 18 August 2026 and GLM-5.3-Flash on 26 August 2026.
• 1M tokens in, 128K out for both. The GLM-5.3 and GLM-5.3-Flash documentation pages each list a 1M token context window and 128K maximum output tokens.
• $1.40 input and $4.40 output per million tokens for GLM-5.3. GLM-5.3-Flash costs $0.15 and $0.50; cached input is $0.26 and $0.03, and cached input storage is listed as limited-time free.
• Open weights under two licences. GLM-5.3 (753B parameters on its card) uses the GLM-5.3 License; GLM-5.3-Flash (320B total, 18B active) uses the MIT License.
• Reasoning is always on. thinking.type accepts only enabled on both models; reasoning_effort takes low, high or max, with max as the default.

What is GLM-5.3?

GLM-5.3 is the flagship of Z.ai's GLM-5 generation. The model page says it uses the same base model as GLM-5.2, with the improvements coming from post-training. Its card is tagged with the glm_moe_dsa architecture and cites the GLM-5 technical report, "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763). GLM-5.3 accepts text only and is aimed at coding, agentic engineering and long-horizon tasks.

GLM-5.3-Flash is a separate, smaller model. Z.ai calls it "the first native multimodal model in the GLM-5 series" and describes it as the first open-source frontier model to combine sparse and linear attention. Its documentation lists:

• Size: 320B total and 18B activated parameters, with 45 layers, pre-trained on a 30T token multimodal corpus.
• Attention: a hybrid of linear attention for local dependencies and sparse attention with a lightweight indexer for global context; an IndexPool step compresses four indexer key vectors into one.
• Efficiency claim: against GLM-5.3, attention compute and KV cache are reduced by about 3.0 and 4.4 times.
• Variants: GLM-5.3-FlashX, served at a stated 200 tokens per second, and GLM-5.3-FlashX-Legal, tuned for legal research, contract review and drafting.

For the full Z.ai line-up, see Zhipu Z.ai GLM Models. For background on sparse expert designs, see Mixture of Experts Explained.

Key facts at a glance

As of 2026-11-15. Source: Z.ai model pages, pricing page, release notes and Hugging Face model cards.

Release and version history

As of 2026-11-15. Dates from Z.ai's release notes.

The release notes carry no dated entries for GLM-5.3-FlashX or GLM-5.3-FlashX-Legal. The Flash page says that before release the model was tested anonymously as "ox-alpha" on OpenCode and OpenRouter, and that it has been served on Chinese AI chips.

How much does GLM-5.3 cost?

As of 2026-11-15. USD per million tokens, as the pricing page states. Source: Z.ai pricing page.

• Same price as GLM-5.2 and GLM-5.1: GLM-5.3 is listed at the same rates as both, so moving between them changes behaviour, not unit cost.
• GLM Coding Plan: GLM-5.3 and GLM-5.3-Flash are fully available on the subscription plan, which uses a points-based quota. GLM-5.3-Flash gets three times the quota of GLM-5.3, and off-peak calls, including all day at weekends, use 50% of standard points. GLM-5.3-FlashX is not yet on the plan.
• Reasoning tokens: because thinking is always on and the default effort is max, reasoning output is billed on every call unless you lower the effort.

Worked example for GLM-5.3: 100,000 input tokens and 10,000 output tokens cost $0.14 for input plus $0.044 for output, about $0.18 in total. On GLM-5.3-Flash the same request costs about $0.02. For how caching changes the maths, see LLM Pricing Explained: Tokens and Caching.

Context window and output limits

Both models list a 1M token context window and 128K maximum output on their Z.ai pages. Neither Hugging Face card states a general context window; they give evaluation settings instead, for example 1M context for SWE-Marathon on GLM-5.3 and 300K context for HLE with tools on both.

Third-party hosts set their own limits, so check the host's documentation. For what context figures mean in practice, see LLM Context Windows Explained.

Modalities and tools

GLM-5.3 accepts text and returns text. GLM-5.3-Flash accepts video, images, text and files and returns text, with images passed by URL (recommended) or Base64 data URL; multiple images are supported.

• Capabilities listed (both): thinking mode, streaming, function calling, context caching and structured output such as JSON.
• Reasoning control: thinking.type takes only enabled; reasoning_effort takes low, high or max (default max), and Z.ai recommends max for complex tasks such as coding. On the Flash page Z.ai recommends thinking.clear_thinking set to false, and stream plus tool_stream for streaming.
• Migration from GLM-5.2: apps that send thinking.type disabled must switch to enabled and set reasoning_effort to low before changing the model ID to glm-5.3, or requests fail.
• Sampling: examples use temperature 1.0; the Flash page also recommends top_p 0.95.
• Endpoints: OpenAI Chat Completion, OpenAI Responses and Anthropic Messages compatible base URLs are listed. Former GLM Coding Plan subscribers, including expired ones, can currently use only the OpenAI Chat Completion protocol.
• SDKs: zai-sdk for Python and Java, and the OpenAI Python SDK.
• Self-hosting: both cards list vLLM, SGLang, TokenSpeed, Transformers, KTransformers, Unsloth and Docker Model Runner, plus quantised builds for llama.cpp, Ollama and LM Studio; the GLM-5.3 card adds vLLM-Ascend, xLLM and SGLang for Ascend NPUs.

Anthropic-compatible endpoints let coding agents switch to GLM by changing a base URL and key. Treat that key as a credential with spend authority. See the AI Coding Agents Security Guide.

Where is GLM-5.3 available?

As of 2026-11-15. Source: Z.ai model pages and Hugging Face model cards.

The Flash page recommends GLM-5.3-FlashX-Legal for use with specialised MCP services and the AutoClaw Legal Assistant, which has Word and WPS plugins.

Access restrictions

GLM-5.3 and GLM-5.3-Flash are not restricted-access models: both are listed on the public pricing page and their weights are downloadable. One licence condition applies to GLM-5.3. The GLM-5.3 License grants broad rights to use the software but requires Model as a Service operators with aggregate revenue above 10 billion US dollars over any 12 consecutive months to pass a Z.ai security review before commercial use. The licence defines MaaS to exclude end-user products that embed model capabilities in specific features and the relaying of requests to models hosted by others. It also requires the copyright and licence notice to be kept and use to comply with applicable laws. GLM-5.3-Flash is under the MIT License with no such condition. For licence questions across open models, see Open-Weight vs Closed LLMs Explained.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported from the GLM-5.3 and GLM-5.3-Flash model cards and Z.ai model pages, published with the August 2026 releases. Comparators shown are Z.ai's own previous model, GLM-5.2, as reported in the same source. Independent leaderboard results are not included because none were checked against their own published pages for this profile.

The GLM-5.3 card states that for the cyber evaluations Git information was removed and a domain allow-list applied, and it says cyber capability "developed faster than we expected" as post-training scaled. The model page also reports 2,436 vulnerabilities found across 269 projects in real-world security work. The cards report scores for models from other providers too; those are Z.ai's own runs and are not reproduced here. For how to read these figures, see How to Read LLM Benchmarks.

Model card and system card

Z.ai publishes Hugging Face model cards for GLM-5.3 and GLM-5.3-Flash and documentation pages on docs.z.ai. It publishes no separate system card for either model on the pages checked. What the cards state:

• Architecture: GLM-5.3 is tagged glm_moe_dsa and shares GLM-5.2's base model; GLM-5.3-Flash uses hybrid linear and sparse attention, IndexPool and Manifold-Constrained Hyper-Connections.
• Evaluation safeguards: for NL2Repo, rule-based and LLM-based checks blocked behaviours such as unauthorised pip or curl operations; for cyber benchmarks, Git information was removed and a domain allow-list applied.
• Safety: neither card has a dedicated safety or limitations section.

With rising cyber scores and no published safety section, run your own evaluation and red-team tests before giving either model tool access. See the Open-Weight Model Security Guide.

Data handling

Z.ai has no model-specific data policy for GLM-5.3. The privacy policy and data processing addendum (last updated 29 September 2025) apply to the API Platform:

• Contracting entity: Jingsheng Hengxing Technology Pte. Ltd. and its affiliates, registered in Singapore.
• Processing location: "Customer Data is generally processed in Singapore."
• Inputs and outputs: the DPA states that content inputs and outputs are not saved on Z.ai's servers; other customer data is stored temporarily to provide the API services and deleted after the terms end.
• Purpose limit: customer data is processed only on the customer's behalf to provide and support the API services. The model-training language in the privacy policy applies to individual users, not API customers.

Self-hosting the open weights keeps prompts in your own environment and shifts the risk to model supply chain and serving security. See the Self-Hosted LLM Security Guide.

Deprecation status

As of 15 November 2026, Z.ai publishes no deprecation or shutdown date for GLM-5.3 or GLM-5.3-Flash. Earlier GLM-5 models remain on the pricing page.

Track dates across providers in the LLM Deprecation and Retirement Tracker.

How GLM-5.3 differs from other models in its family

As of 2026-11-15. Source: Z.ai release notes, pricing page, model pages and model cards. Models listed by release date.

The documented differences between GLM-5.3 and GLM-5.3-Flash are size, price (Flash input and output cost about a ninth of GLM-5.3's), multimodal input, licence and Coding Plan quota. GLM-5.3 differs from GLM-5.2 through post-training, mandatory thinking and its reported benchmark gains, at the same price. For cross-provider fields, see DeepSeek vs Qwen vs Kimi vs GLM.

Practitioner checklist

• Search code for thinking.type disabled and change it to enabled with reasoning_effort low before switching to glm-5.3.
• Set reasoning_effort explicitly; the default of max raises output tokens and cost on every call.
• Choose GLM-5.3-Flash when you need image, video or file input, since GLM-5.3 accepts text only.
• Check the licence for your use: GLM-5.3 carries the MaaS security review condition above the revenue threshold; GLM-5.3-Flash is MIT.
• Confirm the endpoint protocol if you hold an old GLM Coding Plan subscription, which is limited to OpenAI Chat Completion.
• Record that hosted API data is generally processed in Singapore and check that against your data residency rules.
• If self-hosting, pin the Hugging Face revision, verify file hashes and isolate the inference server.
• Store Z.ai API keys in a secrets manager, separate them by environment and alert on unusual spend.
• Run your own evaluation and red-team tests before granting tool access, since neither card has a safety section.

For key handling, see the LLM API Key Security and LLMjacking Guide.

Standards and references

• Z.ai: GLM-5.3 model documentation (accessed 2026-11-15)
• Z.ai: GLM-5.3-Flash model documentation (accessed 2026-11-15)
• Z.ai: Pricing (accessed 2026-11-15)
• Z.ai: Release notes (accessed 2026-11-15)
• Z.ai on Hugging Face: GLM-5.3 model card (accessed 2026-11-15)
• Z.ai on Hugging Face: GLM-5.3-Flash model card (accessed 2026-11-15)
• Z.ai on Hugging Face: GLM-5.3 License (accessed 2026-11-15)
• Z.ai: Privacy policy and data processing addendum (accessed 2026-11-15)

Related NHI Mgmt Group resources: Zhipu Z.ai GLM Models, DeepSeek vs Qwen vs Kimi vs GLM, Frontier AI Models Compared, Open-Weight Model Security Guide, How to Set Spend Limits on LLM API Keys.