GLM-5.3 (also written glm 5.3) costs exactly what GLM-5.2 did: $1.40 per million input tokens, $4.40 output, $0.26 cached input. Z.ai says the two share a base model and that every gain comes from post-training. So the question with this release is not price. It is whether the post-training changed enough to justify switching.
The open weights are at zai-org/GLM-5.3, and Z.ai aims the model at complex software engineering and long-horizon agents.
Specs and Breaking Change
| Item | Value |
|---|---|
| API model ID | glm-5.3 |
| Input | Text only |
| Context / max output | 1M / 128K tokens |
| Reasoning | Always on; reasoning_effort low, high or max (default max) |
| Features | Streaming, function calling, context caching, structured output |
The breaking change, from Z.ai's model guide: reasoning can't be turned off. Clients that send thinking.type: "disabled" must switch to enabled and pick an effort before changing the model ID, or requests fail. Z.ai recommends max for coding.
Three protocols are offered: OpenAI Chat Completions (https://api.z.ai/api/coding/paas/v4), OpenAI Responses (https://api.z.ai/api/v1) and Anthropic Messages (https://api.z.ai/api/anthropic). Accounts that have ever held a GLM Coding Plan, expired or not, are limited to the Chat Completions path for now. The Coding Plan bills in points, at 50% off-peak and on weekends.
Vendor-Reported Improvements
All vendor-reported:
| Benchmark | GLM-5.2 | GLM-5.3 |
|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 |
| DeepSWE v1.1 | 46.2 | 66.9 |
| Agents' Last Exam | 23.8 | 28.5 |
| CyberGym | 77.2% | 84.5% |
| ExploitBench | 24.4% | 54.4% |
Z.ai calls the CyberGym score the best to date, ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%), while on ExploitBench it still trails Mythos 5 (78.0%) and GPT-5.6 Sol (76.5%). On its private Code Bench, max effort went from 23.4% at roughly 96K output tokens per task to 34.5% at roughly 75K, and high effort reached 31.4% at about 50K, which Z.ai says beats Claude Opus 4.8's 29.5% at 120K. Z.ai sums this up as a 50% coding gain and claims open-model SOTA on Terminal Bench 3.0 and Agents' Last Exam (CLI). It also reports 2,436 vulnerabilities disclosed across 269 projects with Chinese security teams. None of the private-bench numbers can be checked from outside.
Human Preference and Usage
Arena's October 15, 2026 snapshot (style control on overall text):
| Entry | Overall | Rating (95% CI) | Votes | Coding | Coding rating |
|---|---|---|---|---|---|
glm-5.3-max |
28 | 1478.5 (1473.3–1483.8) | 22,543 | 38 | 1518.5 (1509.9–1527.1) |
glm-5.3-flash |
41 | 1473.8 (1469.0–1478.7) | 32,925 | 27 | 1524.5 (1517.1–1532.0) |
The cheaper Flash beats Max on coding, though both overlap their neighbors.
→ Track it on AI Rank: GLM-5.3 (max) — Arena rating, rank and 90-day history
Flash wins on traffic too. In OpenRouter's text usage for October 10–16, z-ai/glm-5.3-flash-20260826 was No. 2 with 12.4T tokens (+14%) and z-ai/glm-5.3-20260816 No. 7 with 4.54T (+43%). GLM-5.3-FlashX ranked No. 63 with 186B. Upstream display names sometimes add "(batch)"; these are the IDs AI Rank returns.
→ Live usage: See this week's order in the OpenRouter model usage ranking on AI Rank (tokens, Oct 10–16, 2026 window).
Pricing and License
From the pricing page, USD per 1M tokens: GLM-5.3-Flash is $0.15 / $0.03 cached / $0.50, and GLM-5.3-FlashX is $0.37 / $0.075 / $1.25. Cached-input storage is free for a limited time.
The GLM-5.3 License allows use, modification, fine-tuning, deployment and sale, but a Model-as-a-Service operator whose group revenue tops $10 billion in any 12 consecutive months needs a Z.AI security review first. It isn't MIT or Apache, even though Arena's catalog entry for glm-5.3-flash on AI Rank says MIT; the Hugging Face file is what governs redistribution.
Running GLM-5.3 Locally
The Hugging Face release is about 753B parameters stored mostly in FP8, with a separate GLM-5.3-BF16 repository. GLM-5.3-Flash is also open, at about 321B parameters, and it is a different model: Unsloth's guide says Flash was trained on 30T tokens on a newly trained base with hybrid sparse and linear attention.
Z.ai's model card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth as supported, plus vLLM-Ascend, xLLM and SGLang on Huawei Ascend NPUs. For chat use it recommends passing clear_thinking=true, which defaults to false in the chat template.
Memory needs for Unsloth's dynamic GGUF quants, counting RAM plus VRAM or unified memory:
| Quant | GLM-5.3 | GLM-5.3-Flash |
|---|---|---|
| 1-bit | 223 GB | 100 GB |
| 2-bit | 245 GB | 115 GB |
| 3-bit | 290–360 GB | 128–150 GB |
| 4-bit | 372–475 GB | 162–210 GB |
| 8-bit | 810 GB | 350 GB |
Unsloth says its 2-bit GLM-5.3 build (UD-IQ2_M, 239 GB on disk) works on 256 GB machines such as a Mac Studio or two NVIDIA DGX Sparks, and that 3-bit Flash fits 128 GB devices. Low-bit quants lose accuracy: Unsloth reports about 81% top-1 agreement for 2-bit GLM-5.3 and 87% for 3-bit Flash.
Compare before you choose: rankings move as new votes come in — check the current Arena top-rated models on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 15, 2026; pricing was rechecked and Arena, OpenRouter and local-deployment details were added or refreshed on October 22, 2026. We did not run our own benchmarks for this article.