01What is GLM 5.3 and How It Works
GLM 5.3 is the most powerful open AI model ever released, according to the episode. It comes from a lab called ZAI and launched on August 14th, 2026. The surprising part is that it uses the exact same base model as GLM 5.2.
Same size, same architecture, same pre-training. All the improvement came from post-training.
Post-training is where a model learns how to work, plan, use tools, and check its own answers. ZAI spent about an extra month on post-training. They shifted away from coding puzzles and trained on realistic, messy, multi-step tasks.
Some tasks represent days of work for an experienced engineer. They also built a 'factory' to create these tasks, with judge agents and verifiers that grade without seeing the answer key. This makes the model check itself before finishing.
In this episode, Julian Goldie explains that GLM 5.3 is like the same smart student after a month-long internship doing real jobs. That's why it finishes whole jobs on its own instead of handing back incomplete pieces.
02Key Features and Improvements of GLM 5.3
GLM 5.3 brings several big upgrades you'll feel when you use it. The biggest is long-horizon task handling. It can work for up to 10 hours and 600 turns on a single task without losing the plot.
That's six times better than GLM 5.2 on the Terminal Bencher 3.0 test.
- Long horizon tasks: It stays focused on jobs that take hours, not seconds.
- Self-checking: It tests its own work before telling you it's done.
- Goal mode: Use /goal to make it loop until the task is proven complete.
- Compatibility: You can plug it into other tools like Claude Code or OpenCode.
- Token savings: It uses fewer tokens per task than comparable models.
It also beats top closed models on one independent test. On GDP Val, scored by Artificial Analysis, GLM 5.3 got 1,769. That's higher than Claude Fable 5 at 1,743 and GPT 5.6 Sol at 1,730.
03How to Get Started with GLM 5.3
To use GLM 5.3, you can access it through Zedcode. Zedcode is a coding tool that works like Claude Code or Codex. You can start with a free trial, but note that free tokens run out quickly.
Julian recommends starting there.
- Go to zcode.ai and download the app.
- Start a free trial.
- Open a new task and select a project folder.
- In the model dropdown, choose GLM 5.3 (not GLM 5 turbo or 5.2).
- Set effort level to max for building projects.
- Set permissions to 'full access' if you want it to work without asking.
If you're working on a system where you don't want risks, choose 'ask before changes' to approve edits. But for most builds, full access is easier. You can also work outside a project or connect to a remote workspace via Docker or SSH.
04Using GLM 5.3: Build Websites, Apps, and Trackers
The episode shows several examples of building with GLM 5.3. The first is a one-page website for a dog grooming business. The prompt describes the goal, not the steps: 'Build me a one-page website for a dog grooming business.
The goal is to get visitors to book a call. Include a strong headline, three reasons to choose us, a section that answers common worries, and a contact form. Make it clean, modern, and fast on a phone.
Before you finish, test that the form actually saves submissions.'
It worked for 7 minutes 54 seconds, then showed the step-by-step output. You can preview the site in the browser or open the local file. The result looks clean, not like AI slop.
Another example is a weekly planner. Using goal mode (/goal), you can build an app that opens every Monday, shows seven days, lets you add tasks, mark done, and drag leftovers. It also shows how many tasks you finished this week.
That one took 39 seconds for the first version, then 18 minutes for an improved version.
You can also build a mobile app, like a job tracker for a window cleaning business, and add it to your phone's home screen.
05Scheduled Tasks and Automations with GLM 5.3
GLM 5.3 lets you automate daily or weekly tasks using scheduled tasks in Zedcode. You set up an automation that runs without you. For example, you can create a task to research the latest AI automation news every Monday at 9 a.m.
The model will search, check that content is new, and even build a webpage with the results.
To set it up, go to the automation section, create a scheduled task, write your instructions (or use a voice note via WhisperFlow), pick the schedule, choose a project or 'work outside a project', set permissions to full access, select GLM 5.3, and set effort to max. Then create the task. You can edit it later or do a test run.
One important thing: the app must be open and your computer running. If you shut down or close the app, the scheduled task won't run. Julian uses a Mac Studio that's always on.
You can also run multiple tasks in parallel, like a team of agents.
06Skills, Memory, and Plugins to Enhance GLM 5.3
Skills are saved workflows that you can train the agent on. For example, you can create a skill for SEO content creation. Then you call it with 'use your SEO skill to…'.
You can create, edit, import, and export skills. Skills can be detailed instructions that teach any AI agent how to do something.
Memory is another key feature. You can toggle memory on in Zedcode to save context, but Julian prefers using Obsidian, a free notes app. Obsidian acts as an external memory vault that plugs into all your agents.
You tell GLM 5.3 to access your Obsidian vault for context. It can pull information, like recent activities or keyword ideas, and update the vault for you automatically.
Plugins let you connect to external tools. You can install plugins for GitHub, browser use, iOS simulator, and more. Browser use controls your browser, but Julian says it's not very good with any AI model, so he doesn't teach it in this course.
07Results and Benchmarks: GLM 5.3 vs Other Models
The episode shows several benchmark results. On Terminal Bencher 3.0, GLM 5.3 scored 28.3, up from GLM 5.2's 4.6—about six times better in eight weeks. On Deep SWE, it went from 46.2 to 66.9.
SWE Marathon doubled from 19.4 to 42.5. Automation Bench went from 26.2 to 48.2. Agents Last Exam went from 23.8 to 28.5.
But ZAI ran its own tests, so skepticism is fair. The one independent test, GDP Val, was scored by Artificial Analysis. GLM 5.3 scored 1,769, beating Claude Fable 5 (1,743) and GPT 5.6 Sol (1,730).
That's the strongest result in the release.
However, GLM 5.3 does not beat everything. On ZAI's own hardest internal test, Claude Fable 5 leads 39.5% vs 34.5% at max effort. Also, another open model, Kimi K3 from Moonshop, beats GLM 5.3 on several rows in ZAI's own table, and its weights have been downloadable since late July.
08Tips, Limits, and Mistakes to Avoid with GLM 5.3
Julian shares several tips to get the most from GLM 5.3. First, describe outcomes, not steps. Say 'make a page that convinces a nervous customer to book' instead of 'make a blue header with three boxes'.
The model is trained to plan, so let it plan. Second, always give a verification line, like 'test that every link works before you finish'. That's 10 extra words that double quality because the model was trained on verifiers.
Avoid these mistakes:
- Prompting like a chat model with vague one-liners.
- Leaving effort on low for complex builds.
- Not giving it context from memory files.
- Skipping the verification line.
- Starting over instead of iterating when something is 80% done.
Also, GLM 5.3 does not create images. It's open-weight, but you can't run it locally because it's too big. Use the API cloud-hosted instead.
For token savings, keep conversations in one project to benefit from caching (98% hit rate), and schedule big jobs outside peak China time to get half-rate tokens.
09Who Should Use GLM 5.3 and Why It Matters
GLM 5.3 is for anyone who wants to build websites, apps, or automations without a team. Julian recommends using it as a sub-agent rather than your main agent. He lets Claude orchestrate it for token-heavy tasks, saving his Fable 5 limits.
But it's also great as a standalone tool for beginners because you can try it free.
It matters because it's open, powerful, and cheap. You can start with a free trial. The improvement from GLM 5.2 to 5.3 shows how far open models have come.
It's getting better rapidly, and who knows where it goes next.
For the highest level, you can plug GLM 5.3 into an agentic operating system. Julian showed a mission control dashboard that brings all your agents, memory, and automations into one place. That's the top layer for becoming powerful with AI.
You can even build custom workflows with one-click actions like publishing to WordPress.
