GLM-5.3 is a large language model from Z.ai, the lab formerly known as Zhipu AI, built for complex software engineering, long-horizon agent work and security analysis. Its Terminal-Bench 3.0 score grew sixfold over GLM-5.2 without a single change to the base model, purely through scaled post-training. You pay per token on the API or take a flat GLM Coding Plan subscription. Open weights were announced for Hugging Face once a safety review wraps up.
GLM-5.3 reuses the GLM-5.2 base, a mixture of experts architecture at roughly 743 billion parameters with a 1-million-token context window, and draws every gain from post-training at scale. On Z.ai's internal coding bench, the improvement over GLM-5.2 reaches 50%.
The training recipe swaps short coding drills for simulated workstations, complete with codebases, documentation, testing tools and multi-step assignments, the kind where an agent must spot a bottleneck in an ML training stack, fix it, then prove the speedup. Public numbers back the approach.
Efficiency moved too. At max effort the model completes 34.5% of tasks with around 75,000 output tokens, where GLM-5.2 managed 23.4% while burning 96,000. Artificial Analysis even ranks it level with Kimi K3 at the top of open models.
| Benchmark | GLM-5.2 | GLM-5.3 |
|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 |
| DeepSWE v1.1 | 46.2 | 66.9 |
| CyberGym | 77.2% | 84.5% |
| ExploitBench | 24.4% | 54.4% |
Over 2,400 vulnerabilities across 269 software projects, that is the tally Z.ai reports for its model, including a flaw hiding in code authored four decades ago and a serious weakness reportedly spotted in the Cursor editor.
Here is the twist, nobody planned for this. Z.ai added vulnerability-discovery data expecting sharper reasoning on single bugs; as training scaled, the model started chaining full exploitation plans. It now edges past Claude Mythos 5 and GPT-5.6 Sol on CyberGym while sitting well behind them on ExploitBench (fair warning, the figures come from Z.ai itself and independent replication is still pending).
The API charges $1.40 per million input tokens and $4.40 for output, while the Coding Plan lines up three tiers, Lite at $18 per month, Pro at $80, Max at $168, with a 30% discount on yearly billing. Rates shift fast in this market, so treat Z.ai's official page as the only figure that counts.
Setup takes two moves, you point your agent at Z.ai's OpenAI- or Anthropic-compatible endpoint, swap the model ID, and Claude Code, Cline or OpenCode run GLM-5.3 like any of the coding assistants already on your machine. One migration detail deserves attention, reasoning can no longer be disabled, with three effort levels left to pick from.
No, GLM-5.3 has no free tier. Billing goes through Z.ai's pay-per-token API or the flat GLM Coding Plan, whose entry tier remains one of the cheapest tickets into frontier-class coding agents. Once the announced weights land on Hugging Face, self-hosting will remove per-token costs entirely for teams running their own GPUs.
Not quite yet. At launch the model was API-only, with Z.ai committing to publish the weights on Hugging Face roughly two weeks later, once a safety review ended. No license had been confirmed at that point, whereas GLM-5.2 shipped directly under MIT terms, so check the conditions before planning a local deployment.
Partly. According to Reuters, Z.ai introduced a trusted-access system for the most sensitive functions, keeping defensive security work open while gating higher-risk capabilities. OpenAI's Greg Brockman commented that the model could significantly speed up threats, which explains the deepest risk review Z.ai has run before releasing weights.
Budget and workload decide it. Anthropic's models keep the lead on several hard evaluations such as ExploitBench, yet GLM-5.3 beats them on CyberGym, leads open models on Terminal-Bench 3.0 and costs a fraction per token, which makes long agent sessions far easier on the wallet.
Verdict: Running a coding agent for hours across an entire repository is where GLM-5.3 pays for itself; security teams auditing their own codebases will find an unplanned ally in the bargain.
