GLM-5.3
GLM-5.3 is the main model from Z.ai (Zhipu AI), released on Aug 14, 2026, with 744 billion parameters, 40 billion active per token, and open weights on Hugging Face. On the official API it costs US$ 1.40 per million input tokens and US$ 4.40 per million output, with a 1 million token window, but it accepts text only. In Z.ai's own table it leads 3 of 16 tests, in security, automation and office work, and trails GPT-5.6 Sol and Claude Fable 5 on most coding ones.
GLM-5.3 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Z.ai (Zhipu AI)
- API identifier
glm-5.3- Release
- Aug 14, 2026 on the Z.ai blog; weights on Hugging Face with first upload on Aug 27, 2026
- Weights
- open on Hugging Face and ModelScope, under the GLM-5.3 License
- Size
- 744 billion parameters, 40 billion active (8 of 256 experts per token, plus 1 shared); same base as GLM-5.2
- Input
- US$ 1.40 per million tokens
- Output
- US$ 4.40 per million tokens
- Cache read
- US$ 0.26 per million; cache storage free for a limited time
- Batch
- not published by Z.ai
- Context window
- 1 million tokens, single price
- Max output
- 128 thousand tokens
- Reasoning
- always on; effort low, high or max (default max)
- Input and output
- text only
- Reliable knowledge
- date not published by Z.ai
- Where it runs
- Z.ai API, GLM Coding Plan (from US$ 18 per month), chat at z.ai, Alibaba Cloud Model Studio and, with the weights, SGLang, vLLM or TokenSpeed
Where GLM-5.3 does well, where it falls behind and who it is for
Where it does well
- Finding vulnerabilities: 84.5 on CyberGym, the top score in Z.ai's table, ahead of Claude Fable 5 (83.8) and GPT-5.6 Sol (83.6).
- Office work and automation: 1,769 points on GDPval-AA v2, measured by Artificial Analysis, and 48.2 on AutomationBench, the top scores in the same table.
- Price: US$ 1.40 and US$ 4.40 per million, 30% less on input and 56% less on output than Claude Sonnet 5.5 and GPT-6.1 Sol.
- A license with almost no conditions: the only extra rule reaches anyone selling the model as a service with revenue above US$ 10 billion over 12 months.
- Spends fewer tokens: on Z.ai Code Bench, an internal test, it scores 31.4% with about 50 thousand tokens per task at high effort, against 29.5% for Claude Opus 4.8 with 120 thousand.
Where it falls behind
- Behind the closed leaders in coding: 66.9 on DeepSWE against 72.7 for GPT-5.6 Sol, and 78.1 on FrontierSWE against 88.2 for Claude Fable 5, in Z.ai's own figures.
- Text only: it does not read images or video. For that, Z.ai offers GLM-5.3-Flash, a smaller model on a different base.
- Vulnerable to prompt injection: in a Google chart with Gray Swan data, attacks within 15 attempts succeeded 31.5% of the time against GLM-5.3, against 1.0% for Claude Opus 5.5.
- Mandatory reasoning: thinking cannot be turned off. Anyone who used the non-reasoning mode must move to low effort, or the call fails.
- Cache pricier than closed models: US$ 0.26 per million, against US$ 0.20 for Sonnet 5.5 and US$ 0.10 for GPT-6.1 Sol.
Who it makes sense for
- Companies that want to run a strong model on their own infrastructure: the FP8 weights add up to about 756 GB, close to a third of Qwen3.8-Max at the same precision.
- Coding and automation agents on a tight budget, in Claude Code, OpenCode or ZCode, through the Anthropic compatible API.
- Security review of your own code, to find flaws before an outsider does: it is one of the two highlights of the launch, alongside coding.
- Multi-step office tasks, where it has the top GDPval-AA v2 score in Z.ai's table.
When to pick another model
- Any input with images, scanned PDFs or video: GLM-5.3 accepts text only.
- Agents that read email or web pages without their own safeguards, given the 31.5% result in the Gray Swan injection test.
- Long coding work where every point counts: GPT-5.6 Sol and Claude Fable 5 together lead 10 of the 16 tests in Z.ai's table.
- Short, instant answers: it always reasons, and the default effort is max.
What Z.ai published at launch
GLM-5.3 came out on Aug 14, 2026 on the Z.ai blog, with weights promised two weeks later, after safety evaluation; the first upload to Hugging Face is dated Aug 27, 2026. It is the same base model as GLM-5.2: every gain came from post-training.
Z.ai says it improved 50% over GLM-5.2 on Z.ai Code Bench, its internal test. On public tests, the jump shows on Terminal-Bench 3.0, from 4.6 to 28.3, and on DeepSWE, from 46.2 to 66.9.
The headline of the announcement is security. Working with security teams in China, the model found 2,436 vulnerabilities across 269 projects, according to Z.ai. That helps defense, but it shows the other side too: the weights are open to anyone.
In the 16-test table, GLM-5.3 has the top score on 3; GPT-5.6 Sol on 6, and Claude Fable 5 on 4. The closed rivals already have newer versions, released between Sep 1 and Sep 29, 2026, that are not in it.
| GLM-5.3 Z.ai | Kimi K3 Moonshot | Qwen3.8-Max Alibaba | Fable 5 Anthropic | GPT-5.6 Sol OpenAI | |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 terminal |
88.2 | 88.3 | 86.6 | 88.0 | 88.8 |
| Terminal-Bench 3.0 terminal, long tasks |
28.3 | 17.4 | not published | 33.7 | 34.6 |
| DeepSWE v1.1 long-horizon coding |
66.9 | 67.5 | 56.6 | 69.7 | 72.7 |
| FrontierSWE coding |
78.1 | not published | not published | 88.2 | not published |
| SWE-Marathon v1.1 long coding work |
42.5 | 48.1 | not published | 33.1 | 42.5 |
| CyberGym finding vulnerabilities |
84.5 | 80.0 | 78.5 | 83.8 | 83.6 |
| ExploitBench exploiting vulnerabilities |
54.4 | 32.2 | 28.8 | 78.0 | 76.5 |
| Toolathlon Verified tool use |
73.0 | 76.5 | 72.5 | 74.7 | 74.9 |
| AutomationBench automation |
48.2 | 46.7 | 39.8 | 46.2 | 45.8 |
| HLE with tools hard knowledge |
62.5 | 59.8 | 56.2 | 63.9 | 64.5 |
| GDPval-AA v2 office work, Elo |
1,769 | 1,682 | 1,739 | 1,743 | 1,730 |
Figures from the GLM-5.3 card on Hugging Face, as Z.ai published them, without the percent sign; GLM-5.3 at max effort. The original table also includes GLM-5.2, DeepSeek-V4 Pro and Claude Opus 4.8, which leads SWE-Marathon with 48.8. The Fable 5 results may have used fallbacks. The dot marks the best among the columns shown.
Open weights: what changes for a company
With the weights, the model runs on your own or rented servers: prompts never leave your network, the model does not disappear if the API changes, and fine-tuning stays with you. Z.ai points to SGLang, vLLM, TokenSpeed and Transformers, and cites support for Ascend chips.
It is smaller than its open neighbors: about 756 GB in FP8 and 1.51 TB in BF16, summing the files on Hugging Face, against about 2.50 TB for Qwen3.8-Max in FP8 and 1.56 TB for Kimi K3. Z.ai has not published how many GPUs it recommends.
The GLM-5.3 License allows using, modifying, fine-tuning and selling for free, keeping the copyright notice. The only condition applies to anyone selling the model as a service whose revenue, with affiliates, tops US$ 10 billion over 12 months: that company must pass Z.ai's security review before commercial use. The Kimi K3 license draws that line at US$ 20 million, and the Qwen3.8-Max one at US$ 50 million.
On the official API, the data rules are also favorable on paper:
- The terms say Z.ai does not use content from API businesses and developers to develop or improve its services unless the customer explicitly agrees.
- The data processing addendum says submitted content is processed in real time and is not saved on its servers.
- Processing generally takes place in Singapore, where the company that runs the API is based.
- The terms forbid using the model's outputs to train a competing model.
Pricing: GLM-5.3 on the API against its neighbors
On the API, GLM-5.3 is the cheapest in the table below on both input and output. It costs 30% less than Qwen3.8-Max on input and about 27% less on output, and less than half of Claude Sonnet 5.5 and GPT-6.1 Sol on output. Only the cache read is pricier than on the closed models. No maker has put the five in the same price table; the independent scores are at the top of this page.
| GLM-5.3 Z.ai | Qwen3.8-Max Alibaba | Kimi K3 Moonshot | Sonnet 5.5 Anthropic | GPT-6.1 Sol OpenAI | |
|---|---|---|---|---|---|
| Input | US$ 1.40 | US$ 2 | US$ 3 | US$ 2 | US$ 2 |
| Output | US$ 4.40 | US$ 6 | US$ 15 | US$ 10 | US$ 10 |
| Cache read | US$ 0.26 | US$ 0.25 or US$ 0.17 | US$ 0.30 | US$ 0.20 | US$ 0.10 |
| Context window | 1 million | 1 million | 1,048,576 | 1 million | 1,050,000 |
| Long prompts | single price | single price | single price | single price | double input above 272 thousand |
| Downloadable weights | yes | yes | yes | no | no |
Standard pricing; Qwen3.8-Max is the Singapore region price. Z.ai, Alibaba Cloud, Moonshot, Anthropic and OpenAI documentation checked on Oct 1, 2026.
How to test before adopting
Start with the API, with ten real tasks on GLM-5.3 and on your current model, comparing accuracy and cost per task. Also test high and low effort, which are cheaper than the default max. Add your own prompt injection safeguards before connecting the model to email or the web. If the result holds up and the data cannot leave the building, then price servers for the weights.
Examples of using GLM-5.3
The model runs inside the company network and scans the repository for flaws before deployment. It is one of the two highlights of the launch: 84.5 on CyberGym, the top score in Z.ai's table.
GLM-5.3 plugs into Claude Code through the Anthropic compatible API and works in the repository, with output at US$ 4.40 per million tokens.
Contracts, meeting minutes and spreadsheets exported as text become a report. On GDPval-AA v2, an office work test, it scores 1,769 points in Z.ai's table, the highest in the comparison.
Where GLM-5.3 stands out and where it falls behind, by the numbers
GLM-5.3 has published results on 21 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Nov 17, 2025 and Nov 12, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 21 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 90.9% | #30 / 79 (tied with 3) | #30 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 91.1% | #52 / 81 (tied with 3) | #53 / 168 | 100.0% GPT-5.5 |
| Dtbench | 87.7% | #46 / 68 | #49 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 55.51 | #10 / 68 | #10 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 19.1% | #28 / 78 | #28 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 75.4% | #15 / 52 | #15 / 112 | 93.6% GPT-6 Astra |
| SciCode | 59.0% | #11 / 75 | #11 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 68.8% | #28 / 68 | #28 / 86 | 93.7% GPT-6 Astra |
| WebDev Arena | 1,614 | #17 / 73 | #17 / 82 | 1,820 Claude Opus 5.5 |
| SimpleQA Verified | 41.0% | #40 / 63 (tied with 1) | #45 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 37.0% | #34 / 63 (tied with 1) | #36 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 33.0% | #25 / 61 | #25 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 49.0% | #32 / 70 | #32 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 29.3% | #31 / 58 (tied with 2) | #31 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 8,164 | #12 / 48 | #12 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 56.6% | #16 / 49 | #16 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 40.1% | #25 / 43 | #25 / 43 | 54.6% Claude Opus 5.5 |
| DeepSWE | 69.0% | #7 / 26 | #7 / 26 | 74.1% GPT-6 Astra |
| FrontierSWE | 30.2% | #10 / 20 | #10 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 42.6% | #8 / 16 | #8 / 16 | 57.8% Claude Opus 5.5 |
| Btf3 | 14.9% | #1 / 8 (tied with 1) | #1 / 8 | 14.9% GLM-5.3 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare GLM-5.3 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| Claude Opus 4.7 | Anthropic | 156.32 | 1,532 | 76.5 | Apr 16, 2026 |
| Grok 4.6 | xAI | 156.31 | 1,475 | 78.0 | Aug 12, 2026 |
| GPT-6 Luna | OpenAI | 156.31 | n/a | 72.0 | Sep 22, 2026 |
| GPT-5.2 Pro | OpenAI | 156.03 | n/a | n/a | Dec 11, 2025 |
| GLM-5.3 | Z.ai (Zhipu AI) | 155.91 | n/a | 76.1 | Aug 14, 2026 |
| DeepSeek V4.1 Flash | DeepSeek | 155.57 | n/a | 81.1 | Sep 9, 2026 |
GLM-5.3 or DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is the model from another lab closest to GLM-5.3 in capability. Across the 15 tests both took, this is the score:
| Test | GLM-5.3 | DeepSeek V4.1 Flash |
|---|---|---|
| Apex Agents | 56.6% | 39.5% |
| CritPt | 19.1% | 14.3% |
| Game puzzles | 33.0% | 43.0% |
| Chess puzzles | 37.0% | 48.0% |
| Lmca | 55.51 | 46.97 |
| SciCode | 59.0% | 51.9% |
| ProofBench | 49.0% | 54.0% |
| FrontierCode | 40.1% | 43.9% |
| Frontiermath Tier 4 V2 | 29.3% | 26.8% |
| AIME 2024 and 2025 (OTIS) | 91.1% | 98.3% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What GLM-5.3 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
Other models from Z.ai (Zhipu AI)
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| GLM-5.3-Flash | Aug 20, 2026 | 153.23 | 1,528 | 71.6 | Open weights (unrestricted) |
| GLM-5.3 | Aug 14, 2026 | 155.91 | n/a | 76.1 | API access |
| GLM-5.2 | Jun 16, 2026 | 151.86 | n/a | 73.2 | Open weights (unrestricted) |
| GLM-5.1 | Apr 7, 2026 | 149.99 | 1,514 | n/a | Open weights (unrestricted) |
| GLM-5 | Feb 11, 2026 | 145.78 | 1,511 | n/a | Open weights (unrestricted) |
| GLM-4.7 | Dec 22, 2025 | 143.79 | 1,495 | n/a | Open weights (unrestricted) |
| GLM-4.6 | Sep 30, 2025 | 140.78 | 1,500 | n/a | Open weights (unrestricted) |
Have you used GLM-5.3? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about GLM-5.3
How much does GLM-5.3 cost?
On the Z.ai API, US$ 1.40 per million input tokens, US$ 4.40 per million output and US$ 0.26 per million for cache reads, with cache storage free for a limited time. There is also the GLM Coding Plan, from US$ 18 per month, for use in coding tools. The weights are free.
Is GLM-5.3 open source?
The weights are on Hugging Face and ModelScope under the GLM-5.3 License, which allows using, modifying, fine-tuning and selling without paying Z.ai. The only condition reaches companies selling the model as a service with revenue above US$ 10 billion over 12 months, which must pass a Z.ai security review.
Can I run GLM-5.3 on my own servers?
Yes. The weights add up to about 756 GB in FP8 and 1.51 TB in BF16, and Z.ai points to SGLang, vLLM, TokenSpeed and Transformers. It has not published how many GPUs it recommends.
Is GLM-5.3 better than Claude and GPT?
In Z.ai's own table, it leads 3 of 16 tests, in security, automation and office work, and trails GPT-5.6 Sol and Claude Fable 5 on most coding ones. Newer models, such as Claude Opus 5.5 and GPT-6.1 Sol, are not in the comparison; the independent scores are at the top of this page.
Is my data safe on the Z.ai API?
The terms say Z.ai does not use API customer content to improve its services without explicit consent, and the data addendum says content is processed in real time, without being stored, generally in Singapore. For sensitive data, the safest way out is running the weights on your own infrastructure.
Where does GLM-5.3 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GLM-5.3 ranks #30 out of 267 models, with an ECI of 155.91. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is GLM-5.3 better than DeepSeek V4.1 Flash?
Across the 15 tests both models took and Epoch AI compiled, GLM-5.3 comes out ahead on 8, behind on 7 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is GLM-5.3 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of GLM-5.3 are: Btf3 (#1 of 8); Lmca (#10 of 68); SciCode (#11 of 75). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Z.ai, GLM-5.3 launch post
- Z.ai, GLM-5.3 documentation
- Z.ai, GLM-5.3 card on Hugging Face (benchmarks)
- Z.ai, GLM-5.3 License
- Z.ai, GLM-5 repository on GitHub (size and deployment)
- Z.ai, pricing
- Z.ai, migration guide to GLM-5.3
- Z.ai, release notes
- Z.ai, GLM Coding Plan
- Z.ai, terms of use
- Z.ai, privacy policy and API data processing addendum
- Alibaba Cloud Model Studio, pricing (GLM-5.3 and Qwen3.8-Max)
- Google, Gemini 4 Argon announcement (Gray Swan prompt injection chart)
- Kimi API, pricing
- Moonshot AI, Kimi K3 License
- Qwen, Qwen3.8-Max License
- Anthropic, pricing (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6.1 Sol page (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking