# GLM-5.3

Z.ai, released 2026-08-14

**Verdict: conditional.** GLM-5.3 is Z.ai's open-weights coding and cyber-defense model on the GLM-5.2 base, with every gain from post-training. Pricing is published at $1.40 input / $4.40 output per 1M, the same as GLM-5.2, and the weights are now on Hugging Face. It stays conditional because its headline coding and cyber scores come from Z.ai's own evaluations: reach for defensive security and agentic automation, but audit it yourself before standardizing. Last verified 2026-08-15.

## Pricing

- Input: $1.40 / 1M tokens
- Output: $4.40 / 1M tokens
- Context: 1M ctx; Flash $0.15/$0.50, launch promo ended

## Strengths

- Defensive cybersecurity: CyberGym 84.5%, top published result ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%)
- Long-horizon coding: DeepSWE v1.1 66.9% (from 46.2%), Terminal-Bench 3.0 28.3 (six-fold over 4.6)
- Agentic automation: AutomationBench v1.0.6 jumps 26.2 to 48.2 (+84%); SAO reinforcement learning drives the long-horizon gains
- Cost-sensitive frontier-adjacent coding: re-verified on Z.ai's official table at $1.40 input / $4.40 output per 1M, the same standing rate as GLM-5.2 and GLM-5.1 despite the post-training gains
- Self-hosters: the open weights are now published on Hugging Face (zai-org/GLM-5.3, about 756 GB of safetensors) on the same GLM-5.2 base
- Deep offensive exploitation: ExploitBench 54.4% still trails Mythos 5 (78.0%) and GPT-5.6 Sol (76.5%)
- Budgeting the Flash tier off the launch discount: the 50% promo is no longer on Z.ai's pricing table, which lists GLM-5.3-Flash at $0.15 input / $0.50 output per 1M
- Modest self-hosting hardware: the published checkpoint is roughly 756 GB, a multi-GPU node deployment rather than a workstation one
- Vision and multimodal workloads: no multimodal capability announced or benchmarked for GLM-5.3; Z.ai's vision line is a separate GLM-4.6V family
- Buyers who need independently reproduced results: every headline benchmark remains vendor-run

## Benchmarks

- CyberGym: 84.5% (https://felloai.com/glm-5-3/)
- DeepSWE v1.1: 66.9% (https://felloai.com/glm-5-3/)
- Terminal-Bench 3.0: 28.3 (https://byteiota.com/glm-53-open-weight-coding-emergent-cyber/)
- Terminal-Bench 2.1: 88.2 (https://the-agent-report.com/2026/08/glm-5-3-zai-post-training-coding-cyber/)
- ExploitBench: 54.4% (https://officechai.com/ai/z-ai-releases-glm-5-3-beats-fable-5-and-gpt-5-6-sol-on-cyberbench/)
- ExploitGym (2h / 6h tasks): 105 / 130 (https://officechai.com/ai/z-ai-releases-glm-5-3-beats-fable-5-and-gpt-5-6-sol-on-cyberbench/)

## Sources

- [Digital Applied — GLM-5.3 post-training launch](https://www.digitalapplied.com/blog/glm-5-3-launch-post-training-scaling-coding-agents)
- [Hugging Face: zai-org/GLM-5.3 model card](https://huggingface.co/zai-org/GLM-5.3)
- [FelloAI — GLM 5.3: Benchmarks, Pricing and the Held-Back Weights](https://felloai.com/glm-5-3/)
- [Z.ai pricing](https://docs.z.ai/guides/overview/pricing)
- [OfficeChai — Z.AI Releases GLM 5.3](https://officechai.com/ai/z-ai-releases-glm-5-3-beats-fable-5-and-gpt-5-6-sol-on-cyberbench/)
- [Axios — China's Z.ai holds GLM 5.3 release over hacking risks](https://www.axios.com/2026/08/14/china-open-source-ai-glm-53)
- [Unite.AI — Z.ai Launches GLM-5.3](https://www.unite.ai/z-ai-launches-glm-5-3-with-frontier-coding-and-a-cyber-capability-that-outgrew-its-training/)
- [The Agent Report — GLM-5.3: Post-Training Alone](https://the-agent-report.com/2026/08/glm-5-3-zai-post-training-coding-cyber/)
- [Kingy AI — GLM-5.3 specs and the missing pricing row](https://kingy.ai/blog/glm-5-3-specs-benchmarks-api-how-to-use/)
- [byteiota — GLM-5.3: Open-Weight Coding SOTA and Emergent Cyber Risk](https://byteiota.com/glm-53-open-weight-coding-emergent-cyber/)

Canonical: https://neomanex.com/models/glm-5-3