Snowflake: GLM-5.2 Challenges Opus 4.7 with Lower Costs

Snowflake: GLM-5.2 Challenges Opus 4.7 with Reduced Cost

The CEO of Snowflake believes that GLM-5.2 is competitive with Opus 4.7 at a fraction of the cost.

In a programming benchmark conducted by Snowflake, the Chinese AI model GLM-5.2 and Anthropic's Opus-4.7 achieved nearly identical performance, solving 66 and 67 percent of the problems, respectively, over three attempts per task.

Opus has a slight edge in terms of accuracy on the first attempt, achieving 53.7 percent compared to 47.6 percent for GLM, and is also more efficient overall: GLM requires an average of 99 iterations per task, compared to 80 for Opus, and consumes almost twice as many tokens.

Despite these efficiency gaps, GLM-5.2 is significantly cheaper at $4.40 per million output tokens, creating significant pricing pressure that could challenge the high valuations of Western AI companies like OpenAI.

Snowflake compared GLM-5.2 and Opus 4.7 in a practical benchmark. The Chinese model held its own.

The test covered 103 tasks, each executed three times, where the models had to write code that worked on both DuckDB and Snowflake. When each model had three attempts per task, both were neck and neck: 66 % versus 67 % of tasks solved.

The accuracy on the first attempt diverges: Opus achieved 53.7 %, while GLM only managed 47.6 %, indicating that GLM's output is less consistent. The Chinese model also averaged 99 executions per task compared to 80 for Opus and consumed 860 million tokens, nearly double the 439 million of Opus.

Opus 4.7 is the better model, but GLM is competitive in Snowflake's code benchmark and costs much less.

GLM's strength lies in its ability to reliably validate code on both platforms (DuckDB and Snowflake) simultaneously. According to Snowflake CEO Sridhar Ramaswamy, this is why only GLM was able to solve certain tasks.

Its weaknesses include a tendency to give up too early and obsessively check the wrong things. In one task, GLM made 411 tool calls in 24 minutes, checking line counts, distributions, null values, and column types, and failed all three of its attempts. Opus solved the same task with 49 calls in 9 minutes.

The claim that GLM produces cleaner code did not hold up, Ramaswamy stated. More checks do not lead to more correct results. Despite this, the team is excited about GLM-5.2 and wants to make it available to customers.

China's Pricing Puts Real Pressure on the Western AI Bubble

The results are particularly significant in the context of pricing. GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output tokens, according to Zhipu's official pricing sheet. Some third-party providers offer even lower rates than Zhipu.

• Claude Opus 4.7 costs $5 for input and $25 for output.
• GPT-5.5 costs $5 for input and $30 for output.

GLM's higher token usage somewhat reduces this price gap. However, Anthropic and OpenAI are facing serious pricing pressure, which directly impacts coding, the flagship use case on which both Western AI labs are betting.

If this pressure slows revenue growth, or worse, reduces it, the already inflated AI market faces a real test of resilience. The valuations of OpenAI and Anthropic are based on the assumption that revenues continue to grow rapidly. These valuations are tied to billions of dollars in investments in AI infrastructure development, ranging from data centers to chip orders.