GLM Models: GLM-5.3, 5.3 Flash and 5.2 Compared

GLM is Z.ai's model family. The flagship GLM-5.3 is a text-only open-weight model of about 753B parameters with a 1M-token context, a 128K maximum output and a custom licence. GLM-5.3-Flash is a smaller multimodal MIT model, 320B with 18B active. On the API they cost $1.40 and $4.40, and $0.15 and $0.50, per million tokens. Checked 31 October 2026.

The lineup

Facts are from the Hugging Face cards, config files and Z.ai's developer documentation, read on 31 October 2026.

Z.ai says GLM-5.3 uses the same base model as GLM-5.2 and that "every gain comes from post-training". GLM-5.3-Flash is different: a newly trained base with a hybrid of sparse and linear attention, which Z.ai calls a first for the GLM series, trained on a 30-trillion-token multimodal corpus. Z.ai says Flash outperforms GLM-5.2 across benchmarks at one tenth of the price. That is the maker's claim.

There is also GLM-5.3-FlashX, a variant listed on the pricing page at $0.37 input and $1.25 output, which Z.ai groups with Flash in its docs. Its weight status was not clear:.

Licences

GLM-5.3-Flash and GLM-5.2 are MIT. Nothing further to check.

GLM-5.3 is under the GLM-5.3 licence, which we read. It starts with the broad MIT-style permission to use, copy, modify, deploy and sell, including fine-tuning, and has one restriction. If you or an affiliate run a Model-as-a-Service business (giving third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data), and your group revenue passes 10 billion US dollars over any consecutive 12 months, you must pass a Z.ai security review before using the software commercially. End-user products where the model sits behind specific features, and mere request relaying, are excluded. The scale of that line makes the licence effectively permissive for nearly everyone.

Compare with 20 million dollars in Kimi K3's licence and 50 million in Qwen3.8's flagship. For the whole picture see our open source LLM page.

API prices

From Z.ai's pricing page, per million tokens in US dollars, checked 31 October 2026.

The page marks cached input storage as "Limited-time Free". It also lists web search at $0.01 per use. Two GLM-4 era models, GLM-4.7-Flash and GLM-4.5-Flash, are listed as free.

Put that next to the closed models. GLM-5.3 at $1.40 and $4.40 is below Claude Sonnet 5.5 ($2 and $10), and Flash at $0.15 and $0.50 matches DeepSeek's flash on input and undercuts it on output off-peak ($0.60). See LLM API prices and DeepSeek models.

A worked example: an agent run with 3 million input tokens and 300,000 output tokens costs $4.20 plus $1.32, so $5.52 on GLM-5.3, and $0.45 plus $0.15, so $0.60 on Flash, before cache discounts. Z.ai's own coding-benchmark note adds a useful point: at max effort GLM-5.3 reaches 34.5 percent on its private Z.ai Code Bench at about 75K output tokens per task, against 23.4 percent at 96K for GLM-5.2. If that holds for you, GLM-5.3 is cheaper per solved task than its price-equal predecessor.

The GLM Coding Plan

Z.ai also sells a subscription instead of per-token billing. Its documentation says the GLM Coding Plan can be used with agent tools such as ZCode, Claude Code, Codex, OpenCode, AutoClaw and OpenClaw, and lists 20 or more supported tools including Cline. Plans start at 18 US dollars a month, with Pro and Max tiers above. Key terms from the docs:

• All plans support GLM-5.3 and GLM-5.3-Flash.
• Requests for GLM-5.2 and GLM-5.1 are routed automatically to GLM-5.3, and requests for GLM-4.7 to GLM-5.3-Flash.
• Each plan has a 5-hour credit limit and a weekly limit. The first rows are Lite at 2,000 credits per 5 hours and 10,000 weekly, and Pro at 12,000 and 60,000.
• All plans include vision understanding and web search, web reader and Zread MCP servers.

How many credits a typical task burns, and the Max plan numbers, were not on the page we read:. Note the automatic routing: if you pinned GLM-5.2 for reproducibility, the plan will not honour that.

Published scores

All vendor-reported, from Z.ai's GLM-5.3 card. Most agentic tests ran in Claude Code 2.1.207 at max effort. Z.ai also lists rivals; the rows use the best score available to Z.ai.

On Z.ai's reading, GLM-5.3 is best open-weights on Terminal Bench 3.0 and Agents' Last Exam, and ahead on CyberGym, where it is also ahead of the closed models listed. It trails GPT-5.6 Sol on most rows and Claude Fable 5 on several. We cover the cross-lab picture on best open source LLM for coding, and the older generation on GLM 5.2 benchmarks.

The card makes a claim worth flagging: a cyber capability "developed faster than we expected" in post-training, with GLM-5.3 more than doubling GLM-5.2 on exploitation benchmarks (ExploitGym 105 at two hours against 29, per the card). If your deployment exposes the model to untrusted users, that is a dual-use consideration, not just a feature.

What changed from GLM-5.2 to GLM-5.3

The two share a base model, so the difference is entirely post-training. Z.ai's GLM-5.3 page says the new version carries over the reinforcement learning methods introduced in 5.2, including a technique it calls SAO with compaction, which it credits for keeping gains on long tasks and not only short ones. The headline movements on Z.ai's own table are large:

Z.ai says these pipelines still need a meaningful amount of human-in-the-loop work and that making environment generation and verification more autonomous is a next step. That is a rare note of limitation in a launch document and worth keeping in mind when you read the table.

If you are on GLM-5.2 today, the price is identical ($1.40 and $4.40) and the context is the same, so the upgrade costs nothing in list price. The risk is behavioural: a model that reasons longer or acts differently can break prompts that were tuned to 5.2. Re-run your evaluation set before switching production traffic. On the Coding Plan you have no choice, since requests for 5.2 are routed to 5.3 already.

Effort levels and token efficiency

Reasoning effort is the main cost lever on GLM-5.3. Z.ai's page shows a chart of its private Z.ai Code Bench at different effort levels, with figures in the text that we can quote as vendor-reported.

Read it for the shape, not the numbers. Going from max to high costs about 3 points and saves about a third of the output tokens. If your agent runs thousands of tasks, that trade is worth testing. Z.ai also says GLM-5.3 trails Claude Fable 5 on this test, which is an honest admission from a maker about its own benchmark. Because Z.ai Code Bench is private, no one outside Z.ai can run it.

Flash or the flagship

The practical choice for most teams is between the two 5.3 models.

The Flash card's benchmark chart is an image, so we cannot quote exact scores: . A cheap multimodal model that handles visual coding, such as turning a screenshot into a front end, is the one new thing Flash adds to the lineup. If your agent does not read images and you value the top result, the flagship is the safer pick. If you run high volume or need images, start with Flash and escalate hard tasks to the flagship, which is a routing pattern any gateway can implement. See OpenRouter alternatives for gateways that do this.

Privacy and where data goes

Using Z.ai's API or Coding Plan sends your code to Z.ai. Its retention and training terms were not on the pages we read:. The MIT and custom-MIT weights let you avoid that by self-hosting, which is the point of having them. Using a third-party host moves the question to that host.

Running the weights

The card lists seven serving routes: SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth and Ascend NPU frameworks (vLLM-Ascend, xLLM and SGLang). KTransformers is the usual path for CPU and GPU hybrid setups. Reasoning effort accepts low, high or max, and defaults to max if you pass nothing or an unknown value. For chat use, pass clear_thinking=true explicitly, because the template default is false. Both defaults make an untuned call the slowest, most expensive one, so set them on purpose.

If you only have one GPU, GLM is out of reach; see gpt-oss or Qwen3.8-27B. For rented hardware see Runpod, Modal and Lambda. Mistral's platform lists GLM 5.3 and GLM 5.3 Flash as third-party models, which is one non-Z.ai hosted route. Other hosts:.

Which coding agents can use GLM

Exact base URLs and variables for each are on Z.ai's tool pages:.

Who should pick GLM

Limits and gotchas

• Text only on the flagship. Use Flash for images.
• Default effort is max. Pass low or high explicitly for faster calls.
• Silent routing. Coding Plan requests for older models go to GLM-5.3.
• Self-reported numbers. Z.ai's benchmark tables, including a private in-house test, are not independently reproduced.
• Cyber capability. Worth a security review if untrusted users reach the model.

Our take

GLM-5.3 is the strongest argument that open weights have caught the closed frontier on agentic coding, at least on Z.ai's tables. The licence is gentle, the price is low, and the Coding Plan makes Claude Code with an open model a flat monthly cost. The limits are text-only input on the flagship and numbers that come from the maker. Test it beside DeepSeek and Kimi on your own repository before you commit.

Product names are used descriptively. We are not affiliated with Z.ai.