📚 New to this topic? Read our full guide: What Is Artificial Intelligence?.
Snowflake CEO Says China’s GLM-5.2 Nearly Matches Opus 4.7 at a Fraction of the Cost
The chief of data giant Snowflake just ran a real-world test that should worry expensive AI labs. Snowflake CEO Sridhar Ramaswamy found that a cheap Chinese AI model, GLM-5.2, did coding work almost as well as Anthropic’s pricey Claude Opus 4.7 — while costing far less. He shared the results on June 24, 2026. GLM-5.2 is made by Zhipu, a Chinese AI company also known as Z.ai.
This GLM-5.2 vs Opus 4.7 comparison matters because price is now a huge factor in choosing AI. If a cheaper model gives nearly the same result, businesses will switch. Ramaswamy’s test gives rare, honest numbers from inside a big company.
The test: 103 coding tasks, run three times
Ramaswamy set up a fair test. He gave both models 103 coding tasks. Each task was about writing code for two databases: DuckDB and Snowflake. To be fair, he ran every task three times.
The headline result was close. When the models got three tries, GLM-5.2 solved 66% of tasks and Opus 4.7 solved 67%. That is almost a tie. But on the very first attempt, Opus was clearly stronger: it got 53.7% right, while GLM got 47.6% right. (A benchmark like this just means a fixed set of tasks used to compare models fairly.)
The real story is the price
The big gap is cost. AI models charge by “tokens.” (Tokens are small pieces of text — roughly parts of words — and you pay per million of them, both for what you send in and what the model writes out.) GLM-5.2 is far cheaper per million tokens.
Benchmarks & specs: GLM-5.2 vs Opus 4.7
| Metric (as reported) | GLM-5.2 (Zhipu) | Claude Opus 4.7 | GPT-5.5 |
|---|---|---|---|
| Accuracy with 3 tries | 66% | 67% | — |
| First-attempt accuracy | 47.6% | 53.7% | — |
| Input price (per 1M tokens) | $1.40 | $5.00 | $5.00 |
| Output price (per 1M tokens) | $4.40 | $25.00 | $30.00 |
| Avg. runs per task | 99 | 80 | — |
| Total tokens used in test | 860 million | 439 million | — |
Where GLM-5.2 falls short
Cheaper does not mean perfect. Ramaswamy praised GLM for reliably checking code across both databases. But he flagged real weaknesses. The model sometimes gave up on tasks too early. It also made too many tool calls. (A tool call is when the AI uses an outside tool, like running code, to help solve a task.)
One example was striking. On a single task, GLM made 411 tool calls over 24 minutes and still failed. Opus solved the same task with just 49 calls in 9 minutes. So GLM can be slower and messier, even when the final score is close.
Why “open-weight” Chinese models are a big deal
GLM-5.2 is part of a wave of strong, low-cost models coming out of China. Many of them are “open-weight.” (Open-weight means the company shares the trained model so others can run it themselves, often more cheaply.) This is the opposite of “closed” US models like Opus, which you can only use through the maker’s service.
This shift is changing the whole market. For years, the best AI came almost only from a few US labs at premium prices. Now Chinese rivals are catching up fast on quality while charging far less. That forces every AI company to defend its price. It also gives buyers real choice for the first time.
Ramaswamy’s test is valuable because it is honest and detailed. He did not just trust a marketing chart. He ran the same hard tasks on both models and measured accuracy, speed, token use and cost. That kind of real-world testing is exactly what businesses should copy before they pick a model.
FAQ
Who makes GLM-5.2?
It is made by Zhipu, a Chinese AI company also known as Z.ai.
Is GLM-5.2 better than Claude Opus 4.7?
Not quite. In this coding test, Opus 4.7 was slightly more accurate and far more efficient. But GLM-5.2 came very close while costing much less.
How much cheaper is GLM-5.2?
GLM-5.2 charges $1.40 input and $4.40 output per million tokens. Opus 4.7 charges $5.00 input and $25.00 output. That makes GLM several times cheaper per token.
Why it matters (especially for India and founders)
For Indian startups, cost is often the deciding factor. A model that is “good enough” and 5x cheaper can change the math for a small team. It means AI features that once looked too expensive may now be affordable. This is part of a wider trend where cheaper rivals are crowding the market once ruled by US labs, even as American firms fight to control how their top AI models spread. Founders should test cheaper models on their own real tasks before committing budget.
The takeaway: the gap between costly Western models and cheap Chinese open models is shrinking fast. Accuracy is close; the price is not. For anyone building on AI, the smart move is to test more than one model and pay only for the quality you truly need.
Source: The Decoder.
