Clone ChatGPT Agent with a single sentence? ZhiPu GLM-4.5 first test: zero configuration, full functionality
Today, we bring you the first test of the ZhiPu GLM-4.5 model.
ZhiPu has released its latest GLM-4.5, with a total parameter count of 335B and an activation parameter count of 32B. GLM-4.5-Air: total parameter count of 106B, activation parameter count of 12B.
It demonstrates exceptional performance in reasoning, code generation, and agent capabilities, supports mixed reasoning, and offers excellent cost-effectiveness at 0.8 yuan per million tokens and 2 yuan per million tokens for output, with the high-speed version achieving over 100 tokens per second.
I tested GLM-4.5 in advance and found that although its total parameters are not as high as others, its code capabilities are no less impressive and even better.
One notable feature is that due to the presence of hybrid reasoning, even if your prompt is very short, it can still fill in the necessary information for the page and provide excellent results. If your prompt is detailed, it will follow the prompt very well and make intelligent judgments in determining the coding path.
Moreover, they have integrated various agent capabilities into a single API, such as knowledge base retrieval and search capabilities, making product development much more convenient.
You can even use a single API key and prompt to create a simplified ChatGPT agent mode, and the generated PowerPoint presentations look better than those from GPT.
Of course, they have also made it compatible with Claude Code and can directly replace the Claude Code model. There is a rough tutorial later on.
Let's look at some examples, starting from simple to difficult.
All of the test results below are API outputs and do not use Z.ai's full-stack development mode, which will be even more powerful in the future.
This time, GLM-4.5 can also complete code writing tasks well without complex instructions, such as “Help me write a Gmail email page” or “Help me generate a 3D abstract art piece.”
As you can see, I only asked it to generate a Gmail email page without specifying any details, and it handled it perfectly. All the necessary elements are present, including the ability to click on emails to view details in a separate panel—no shortcuts taken.
As you can see, it automatically applied the Three.js library to complete the task, and the generated results are highly detailed.
Other models either only include one or two elements or the elements don't move on their own, but every element generated by GLM-4.5 is in motion and interacts with others.
Let's try a basic front-end component—our complex schedule component, which supports switching between monthly, weekly, and daily views, as well as creating and editing schedules.
Create a fully functional calendar component that not only displays dates but can also be used to manage events and schedules.
Core features:
Multiple views: Supports switching between month, week, and day views.
Event display: Clearly display events on the calendar. Long events can span multiple days or time slots. Different types of events can have different colors.
Event operations:
Create: Click on a date or time slot to quickly create a new event and pop up a form to fill in the details (title, time, description, participants, etc.).
Edit/delete: Click on an existing event to view details or edit/delete it.
Drag and drop: Support dragging and dropping events to quickly change their date or time.
Adjust duration: Support dragging the edges of an event to adjust its duration.
Take a look at the result. There are no issues, and edit and delete icons have been added to the cards in the schedule for quick access.
The small component is complete. It's been half a year, and everyone should have finished their KPI or OKR reviews.
I'm considering creating a complex OKR management tool that supports creating and modifying progress, dashboards for statistics, and switching between Chinese and English.
Create “OKR Manager”:
Functional Requirements
- Objectives and Key Results
- O: Title, cycle (quarter), person responsible, weight
- KR: Title, type (numerical/percentage/milestone), initial/target/current value, weight
- Automatically calculate the aggregate progress of O (based on KR weight)
- Check-in
- Weekly reporting: progress value, blockers, next week's plan; timeline display
- Dashboard
- Filter by person/team/quarter; radar chart or bar chart to display progress distribution
- Multi-language
- Chinese/English switching (i18n dictionary in the frontend)
Technical Requirements
- HTML + JS + Tailwind + Heroicons
- Charts: Chart.js (CDN)
- localStorage; export JSON
- Progress calculation and input validation (0–100% boundaries)
This is absolutely perfect, friends. He has perfectly fulfilled all requirements and succeeded on the first try with no bugs.
The UI and interaction design are excellent.
For example, icons are used to display time, people, and weights for goals and key results, and progress bars are used to show overall progress.
The progress bar also shows the approximate status at the end, with green text for normal and yellow text for risk. The goals and key results use nested card components.
The layout of the entire information and the key interaction positions are very intuitive.
These interfaces are relatively simple, as they are B-end systems. Now, let's look at a scenario with high visual requirements for C-end users.
Here, I asked him to design the entire order checkout process for a fashion brand e-commerce website, including the shopping cart, shipping information, payment information, and checkout confirmation pages.
Roles and Objectives
—
Served as Senior UX/UI Designer. Created high-fidelity desktop web pages for the “VELLORA” online store (luxurious yet accessible fashion and accessories). Included: a refined shopping cart page and a separate checkout process (3 steps:Shipping • Payment • Review/Confirmation).
【Quiet Luxury · Graphite Neutral】
——
Color Palette (Hex)
- Page Background: #F7F7F7 (Light Gray); Content Cards: #FFFFFF
- Primary Text: #222426; Secondary Text: #6B6E73; Emphasis (Near Black): #0E0E0F
- Brand Accent (choose one and use consistently across the site): #9AA18E (Sage) or #8E7C6D (Mocha)
- Separator/Border: #E7E7E7 (1px thin line); hairline can use rgba(0,0,0,.06)
Font
- H1/H2: Elegant serif Newsreader (Alternative: Cormorant Garamond)
- UI text: Geometric sans-serif Manrope (Alternative: Inter)
- Numbers/prices may use monospace style (Manrope Tabular)
Rounded Corners and Shadows
- Rounded Corners: Buttons and Inputs 12px; Cards/Modals 16px
- Shadows: s-sm: 0 1px 2px rgba(17,17,19,.06); s-md: 0 8px 24px rgba(17,17,19,.08) (used for overlay/modals)
- Separated by 1px lines with a few shadows for decoration, overall restrained
Screen and key layout (keep original functionality/flow, create high-fidelity based on the following structure)
——
1) Shopping cart interface (desktop & mobile)
- Desktop layout: · Left column: Shopping cart item list (long table omitted here).Each product card includes: thumbnail, product name, color/size, unit price, quantity step (+/–), “Save for later,” remove (×) · Right column (at the top): order summary card (subtotal, estimated shipping, taxes/duties, promo code input and verification, total), main CTA “Proceed to checkout,” supports “Continue shopping” secondary link
- Mobile layout: · Vertical scrolling list; summary cards fixed at the bottom (within the safe zone), displaying total and main CTA · Quantity increments and remove operations performed in place to avoid page jumps · Discount code folded, click to expand and enter
2) Checkout process (3 steps) — Shipping • Payment • Review/Confirmation
- General · Top step bar (currently highlighted, completed steps checked, clickable to return and edit) · Form groups have clear separating titles and explanatory text; errors displayed locally · Supports “Return to Cart” and “Continue to Next Step” buttons (primary/secondary levels clearly defined)
- Step 1: Shipping · Fields: Recipient, phone number, email, country/region (linked to province/city/district), address 1/2, postal code; invoice and remarks (optional) · Shipping method card: Standard/Express/Same Day (price and estimated arrival time), with real-time updates to the summary after selection
- Step 2: Payment
· Methods: Credit Card, Alipay/WeChat; card information masked in real time; billing address same as shipping address (checkbox)
· Security and Compliance Notice (small text)
- Step 3: Review & Confirm · Summary: Shipping information, delivery method, payment method last digits, product list and total amount; can be edited on-site and returned to the corresponding step · Agree to Terms (checkbox);Place order CTA; order number and next steps displayed after placing order This is quite complex, as not only are there style requirements, but the content of each page is also very detailed, and this is a complex process involving four pages.
As you can see, apart from the horizontal line in the step bar being slightly misaligned, GLM has done a perfect job.
First is the shopping cart page, where you can select and delete items, and the price and SKU are displayed correctly.Below, they have added a trust endorsement section to increase trust.
They even used a clever design for the discount code, using text links in the main color instead of buttons. Learn something good, GLM, haha.
Finally, the information display and layout of the order page are also perfect, with very clear information hierarchy. The choice of cards and dividing lines is especially well done, summarizing all the information entered earlier, and adding a modification button for the delivery information.
After the order is successfully placed, in addition to the Tost prompt in the upper right corner, a full-screen pop-up prompt is added to the webpage. In terms of interaction, there's nothing to complain about.
As I mentioned earlier, in addition to the powerful model itself, ZhiPu's GLM-4.5 also has a bunch of other Agent tools built-in, allowing you to call them all at once. In addition to API calls, they also support MCP calls, and it's a SEE solution, which is very convenient.
I used an API key to call the GLM-4.5 model to generate a webpage for me. The webpage information comes from ZhiPu's search tool, essentially achieving a simple Manus or ChatGPT agent mode with just a prompt.
Help me call ZhiPu's search to retrieve information about the weather in Shanghai and all information about WAIC, and generate a WAIC conference guide as an HTML dynamic webpage. The webpage generation requirements are:
Use Bento Grid style visual design, with a pure black background and Tesla red #E31937 color as the highlight.
Emphasize large fonts or numbers to highlight key points, with large visual elements in the image to emphasize the focus, creating a contrast with the small elements.
Use a mix of Chinese and English, with large bold Chinese characters and small English characters as accents.
Use simple line graphics for data visualization or as accompanying images
.Use the transparency gradient of the highlight color to create a technological feel, but do not blend different highlight colors together.
Mimic the animation effect on Apple's official website, with the animation triggered when scrolling down with the mouse.
Data can be referenced from online chart components, but the style must match the theme
Use Framer Motion (via CDN)
Use HTML5, TailwindCSS 3.0+ (via CDN), and necessary JavaScript
Use professional icon libraries such as Font Awesome or Material Icons (via CDN)
Avoid using emojis as primary icons.
Do not omit key points of content
WAIC is trending right now, so I asked him to search for related information and create a conference guide page.
His search quality was impressive. Since WAIC information is highly time-sensitive, it is difficult to search for. I have used many overseas search APIs, but they could not find much information and often mixed last year's information with this year's.
However, GLM-4.5 returned comprehensive and highly accurate results covering venues, schedules, highlights, transportation, and more.
The webpage was also well-structured with no layout or display issues.
Additionally, GLM-4.5 supports integration with Claude Code via an Anthropic API endpoint, and the integration process is straightforward.
First, you need a ZhiPu API, which can be obtained from the ZhiPu Open Platform (https://bigmodel.cn/usercenter/proj-mgmt/apikeys)
If you haven't installed Claude Code, you'll need to install it first. After installation, you don't need to start it; just enter the following command in the terminal, making sure to replace the API with your own.
export ANTHROPIC_BASE_URL=https://open.bigmodel.cn/api/anthropic
export ANTHROPIC_AUTH_TOKEN="e6c167e0201240d492572fce4e7230ed.Ld8J4nZd2cnz1uxH"
Then select the first option. When you see the API address “bigmodel” on the Claude Code welcome interface, you're all set. Isn't that simple?
Note that this is only a temporary change. If you want to make it permanent, you can modify .bashrc or .zshrc, so that GLM-4.5 will be used every time you start up. You can discuss this with the AI, and it will be a great teacher.
Additionally, they're offering a “$50 monthly subscription for unlimited use of GLM-4.5” promotion. For just $50, you can enjoy unlimited use! Scan the QR code to sign up—spots are limited, so act fast!
If you're not a developer, you can also select the GLM-4.5 model on Z.ai, which comes with many built-in Agent tools for better results, such as full-stack development and PowerPoint generation.
That's all for today's testing.
As I write this, I suddenly realize that the phrase “late but first” is not an adjective describing GLM-4.5, but a new integration strategy—it turns ‘late’ into “free admission.”
Others spent two years polishing reasoning, code, agents, search, MCP, and Claude-Code compatibility into six separate islands, but ZhiPu simply waited for them all to mature and then welded them together into an aircraft carrier in one go.So when users board the ship, they don't see six betas,
but a patched-up official version: the price is set directly at the lowest standard of the later entrants, but the functionality conveniently fills in all the potholes that their predecessors fell into.
What's even stronger is that it brings stronger capabilities at a lower cost.
At a deeper level, this high level of integration is a reaction against the “fragmented AI era.”
Over the past year, we have become accustomed to breaking tasks into Model A for coding, Model B for search, Model C for prototyping, and then writing glue scripts to bind them together.
GLM-4.5 uses a single API to eliminate the glue itself, reducing “cross-model orchestration” to a simple natural language command.
“Latecomers overtaking early adopters” is not simply a matter of taking a shortcut, but rather turning “integration” itself into a weapon: compressing the path others have taken into a shortcut and packaging scattered innovations into a single upgrade.
Top comments (0)