Z.ai GLM model, released August 14, 2026
GLM-5.3: Z.ai's flagship for coding and agents
Released August 14, 2026, with weights on Hugging Face from August 28: the GLM-5.2 base model with far more post-training, a 1M-token context, and a new license. The specs, Z.ai's benchmarks, today's prices, and what the license means.
GLM-5.3 is Z.ai's current flagship. It uses the same base model as GLM-5.2, and Z.ai says every gain comes from post-training: GLM-5.3 beats GLM-5.2 on every benchmark in Z.ai's comparison, from Terminal Bench 2.1 (88.2 against 81.0) and DeepSWE (66.9 against 46.2) to CyberGym, where its 84.5 is the top score in the table. Z.ai calls it the most capable open-weights model for coding.
It keeps the 1M-token context and adds a reasoning_effort control with three levels: low, high, and max, the default. Z.ai also reports a cyber capability that grew faster than it expected: on exploitation benchmarks GLM-5.3 more than doubles GLM-5.2. The name covers a small family: GLM-5.3 Prime is a faster-serving version of the same model, and GLM-5.3 Flash is a separate, smaller multimodal model under MIT, the one that ran anonymously as Ox Alpha.
An independent reference, not affiliated with Z.ai. Checked against Z.ai's model card, docs and OpenRouter on October 4, 2026.
1M
Context window (tokens)
88.2
Terminal Bench 2.1 (Z.ai)
84.5
CyberGym, top of Z.ai's table
Open
Weights, GLM-5.3 License
Where GLM-5.3 stands today
Specifications
The GLM-5.3 spec sheet
From Z.ai's model card and the published config, with each row linked to its source.
What changed
GLM-5.3 against GLM-5.2
Much stronger long-horizon coding
Terminal Bench 3.0 rises from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, SWE-Marathon from 19.4 to 42.5, and Z.ai reports a 50% gain on its in-house Z.ai Code Bench.
A jump in security work
CyberGym 77.2 to 84.5, and exploitation benchmarks more than double: ExploitGym 29 to 105 solved in two hours, ExploitBench 24.4 to 54.4.
Better at automation and tools
AutomationBench 26.2 to 48.2, Toolathlon Verified 59.9 to 73.0, HLE with tools 54.7 to 62.5.
Same base, same size, same context
No new pre-training: the 1M-token context and the parameter count carry over from GLM-5.2.
A new license
GLM-5.2 is plain MIT; GLM-5.3 adds the $10B model-as-a-service security-review condition described above.
Benchmarks
GLM-5.3 benchmarks, as Z.ai published them
Every score and every comparison model comes from Z.ai's own model card. These are vendor-reported results: useful for direction, not independent proof.
| Benchmark | GLM-5.3 | GLM-5.2 | Kimi K3 | DeepSeek-V4 Pro-0813 | Qwen3.8-Max | Opus 4.8 | Fable 5 (w/ fallback) | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|---|
| Coding | ||||||||
| Terminal Bench 2.1 | 88.2 | 81.0 | 88.3 | 87.9 | ||||
Where to run it
How to use GLM-5.3
Call it through Z.ai's API or a third-party host with any OpenAI-compatible SDK, use it inside a coding agent, or download the weights and serve them yourself.
- Z.ai API
Pricing
What GLM-5.3 costs on every host
On Z.ai's own API it costs $1.40 per 1M input tokens and $4.40 per 1M output tokens (cached input $0.26). 34 hosts list it on OpenRouter, and 22 of them charge less than Z.ai. Sorted cheapest first by blended price (three parts input to one part output).
| Host | Input /1M | Output /1M | Cached input /1M | Precision | Context | Max output |
|---|---|---|---|---|---|---|
Wafer | $0.039 | $3.39 | $0.038 | not stated | 1,048,576 | 943,718 |
Wafer | $0.17 | $3.39 | $0.16 | not stated | 1,048,576 | 943,718 |
Sail Research | $0.20 | $3.40 | $0.15 | fp8 | 1,048,576 | 943,718 |
DeepInfra | $0.5625 | |||||
The GLM family
GLM-5.3 next to the other GLM models
Every GLM model with its own page here, newest first, with Z.ai's current price per 1M input / output tokens.
| Model | Released | Parameters | Context | License | Z.ai price (in / out) |
|---|---|---|---|---|---|
GLM-5.3 Flash Cheap, fast, 1M-context multimodal (the model behind Ox Alpha) | Aug 2026 | 320B total / 18B active | 1M | MIT | $0.15 / $0.50 |
GLM-5.3 Z.ai's current flagship: the strongest GLM for coding and agents, 1M context | Aug 2026 | 753B total (GLM-5.2 base) | 1M | GLM-5.3 License | $1.40 / $4.40 |
GLM-5.2 The newest GLM flagship under plain MIT: 1M context, long-horizon coding | Jun 2026 | 744B / 40B active | 1M | MIT | $1.40 / $4.40 |
FAQ
GLM-5.3, answered
What is GLM-5.3?
Is GLM-5.3 open source?
How much does GLM-5.3 cost?
What is the difference between GLM-5.3 and GLM-5.3 Flash?
What is GLM-5.3 Prime?
How do I control GLM-5.3's thinking?
Build something with GLM-5.3
You have seen the specs, the scores and the price. Turning a model into a product people use is the harder part, and it is the part you can skip wiring up yourself.