GLM-5.2 is the free AI model from Beijing-based Z.ai that Silicon Valley has been talking about all week, especially for coding and “agentic” tasks. Business Insider’s test found the headline reads like a victory lap, but the lived experience is a little more chaotic: the model is often slow, it frequently runs into capacity issues, and at one point it took more than 15 minutes just to get unstuck.

That reliability tradeoff matters because GLM-5.2 is competing in the same mindshare space as frontier paid systems. Z.ai says it supports a 1 million-token context window, which in plain English means it can ingest extremely long text at once, “enough to process hundreds of thousands of words at once,” putting it in the same league as OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8. The catch is that the model’s raw capability can be overshadowed by access constraints, and patience becomes part of the product.

So what happens when you actually use it? The outlet started with a simple writing task: an outreach email for Business Insider seeking interviews with career coaches. After waiting several minutes, GLM-5.2 produced an email that “covered all the essentials” and closely matched the style the writer would typically produce. The critique was mostly logistical, not intellectual. In other words, the model looked competent on the basics, but you had to wait long enough for those basics to arrive.

The next test was more practical: recommending wet cat food for a cat with a sensitive stomach. After another capacity delay, the AI suggested several well-known commercial brands, a prescription option, and general advice for choosing food for sensitive cats. The output included generic guidance plus specific brand names, and while it is the kind of info that tends to align with a veterinarian’s general counsel, the shopping experience itself was imperfect. GLM does not provide direct shopping links for the products it recommends. The workaround was straightforward, but for users who expect a “click to buy” flow, it adds friction.

Trip planning exposed a different kind of gap: not whether the model can think, but whether it can follow every constraint. The test asked for a weekend trip for two from Oakland to Monterey, including hiking, scenic photography spots, antique shopping, restaurants, and a budget hotel. The itinerary was thoughtful and detailed, recommending destinations such as Carmel and Moss Landing, and it accounted for traffic and reservations. The lodging piece was the weak spot. Despite the explicit request for a budget hotel, the model initially skipped that part. After being asked again, it provided several motels, but the prices were not realistic for current rates. One recommendation, Super 8 by Wyndham Monterey, was listed at roughly $100-$150 per night, while current rates are well above $300. That’s a classic AI failure mode: the plan can look coherent while one critical variable is off.

The most revealing test was the design workflow. The writer uploaded a photo of an Art Deco-style amethyst ring and asked GLM-5.2 to create an advertisement for a fictional jewelry business. The newest model, GLM-5.2, spent more than 15 minutes stuck at capacity, so the tester switched to the older 4.7 version. That older model processed the English prompt in Chinese, producing both its reasoning and the finished poster in Chinese, even though the poster language was not requested. Then, a usability issue compounded the confusion: the outcome did not come with a PDF or JPEG version that could be downloaded. When the user asked for regeneration in English, the image disappeared and the underlying HTML broke entirely.

Eventually, capacity opened up and GLM-5.2 took a different approach. Instead of stalling, it guided the tester through an interactive design process with style and color options before generating the final poster. The design itself was not described as polished; it looked more like a bar menu than a luxury jewelry advertisement. But it did function properly and allowed the result to be downloaded. That contrast is the real story for executives: the model can be capable, then stumble on reliability and workflow edges, and then recover when conditions allow.

The bottom line from the test is blunt. GLM-5.2 does not yet match the polish or reliability of the best paid AI models. Capacity limits are frustrating, responses can be slow, and some features have obvious flaws, especially live pricing and design generation. Still, for a free, open-source model, it is “surprisingly capable.” On everyday tasks like writing, research-like information gathering, shopping advice, and trip planning, it can deliver information comparable to what users might get from much more expensive competitors. For peers watching the market, the second-order implication is clear: free models can pressure pricing and adoption, but enterprises will still demand reliability, predictable outputs, and clean toolchains. If Z.ai improves reliability and reduces wait times, GLM-5.2 could become a compelling alternative for users who do not want to pay premium AI subscriptions.