GLM-5.3 is Z.ai's flagship model for complex software engineering and long-horizon agent tasks. It reaches state-of-the-art results among open-weight models on Terminal-Bench 3.0 and Agents' Last Exam, and it does so while using fewer output tokens per task than several closed frontier models. GLM-5.3 is now available on SambaCloud at over 200 tokens/s under the model string GLM-5.3. Read Z.ai's full GLM-5.3 announcement here.
Why GLM-5.3
GLM-5.3 is built for the work that agents actually do: long sequences of dependent steps in a real codebase, terminal, or enterprise system. It scores 28.3 on Terminal-Bench 3.0 and 66.9 on DeepSWE v1.1, state-of-the-art among open-weight models, and it reaches those numbers with fewer output tokens per task than several closed frontier models.
Three things make it a strong fit for teams running agents in production:
- Long-horizon coding and agent work. GLM-5.3 was trained on environments that mirror multi-day expert engineering tasks, with access to compute clusters, codebases, documentation, and experiment results. It is built to own a task end to end rather than wait for a user to decompose it.
- Token efficiency. On Z.ai's internal Code Bench, GLM-5.3 reaches 31.4% at around 50K output tokens per task, ahead of Claude Opus 4.8 at 29.5% with 120K tokens. In agent loops where every step is a model call, fewer tokens per step compounds into lower cost and shorter runs.
- Defensive security. GLM-5.3 holds the best published score on CyberGym, a vulnerability-discovery benchmark, and the US Center for AI Standards and Innovation rated it the most advanced open-weight model for cybersecurity. Teams doing code auditing and vulnerability research get a frontier-class open model for the work.
Where GLM-5.3 is already running
On OpenRouter, the five apps routing the most tokens to GLM-5.3 are all agent harnesses: Claude Code leads at 2.5T tokens, followed by Hermes Agent (895B), Strix (826B), pi (681B), and omp (476B). Three of those are coding agents, and Strix is an open-source security agent that uses GLM-5.3 to find and validate vulnerabilities. Developers today are not only using GLM-5.3 for chat; they are putting it inside long-running loops where every step is a model call, which is exactly the workload where speed matters most.
Here is a trace from a typical agent run on SambaCloud: 41 model calls over 17 minutes and 30 seconds, 5.4 million input tokens, and just under 42,000 output tokens. Each turn re-sends everything the agent has seen so far, so input grows from 38K tokens on the first call to over 114K by the eighth, while output stays small: a few hundred tokens of reasoning and a tool call. Over 91% of that input was served from cache.
Agentic workloads like this are what GLM-5.3 was trained for, and it is why per-call latency matters more than any single benchmark. Running at over 200 tokens/s on SambaCloud, GLM-5.3 keeps each of those steps short enough that the whole run finishes while the developer is still watching.
Get Started
Call it from Python
With a few lines of Python, you can hand GLM-5.3 a real engineering task and let it plan, implement, and review its own work.
GLM-5.3 brings frontier agentic coding and the strongest open-weight security capability to production. SambaNova runs it at over 200 tokens/s.
Try it in ZCode
ZCode is Z.ai's agentic development environment, built for the same long-horizon tasks GLM-5.3 was trained on. It accepts any OpenAI- or Anthropic-compatible provider, so pointing it at SambaCloud takes four steps:
- Open Settings in the lower-left corner of ZCode and go to Model Providers.
- Add a provider and set the OpenAI Base URL to https://api.sambanova.ai/v1 and the Anthropic Base URL to [SambaCloud Anthropic Messages endpoint].
- Paste your SambaCloud API key.
- Save, then pick GLM-5.3 from the model menu and set reasoning effort to max.
From there ZCode runs GLM-5.3 on SambaCloud for every task: understanding a project, writing code, resolving issues, and multi-step workflows that span hours.
Other coding agents
The same base-URL swap works in Claude Code, OpenCode, and Codex CLI. For Codex, use the Responses API endpoint and declare low, high, and max as the supported reasoning levels in your model catalog.
Try GLM-5.3 in the SambaCloud playground, or generate an API key and connect it to your agent harness today.
