Low-cost Chinese AI rivals Anthropic’s flagship model, threatening Silicon Valley’s pricing power

The global race for artificial intelligence supremacy has entered a highly disruptive phase, characterized by aggressive price wars and rapidly narrowing capability gaps between closed-source Western frontier models and open-weight alternatives. A stark illustration of this shift emerged recently when Snowflake Chief Executive Officer Sridhar Ramaswamy shared a detailed analysis comparing Zhipu AI’s newly released GLM-5.2 model with Anthropic’s flagship Claude Opus 4.7. Evaluated on a rigorous, real-world data engineering benchmark, the Chinese open-weight model performed nearly neck-and-neck with Anthropic's top-tier offering[1]. While the evaluation highlighted notable efficiency bottlenecks in GLM-5.2, its dramatic cost advantage—priced at only a fraction of the cost of Claude Opus 4.7—is generating intense pricing pressure[2]. This development could fundamentally alter the economics of the enterprise software market and challenge the multi-billion-dollar valuations of leading Silicon Valley AI laboratories[2].

The detailed evaluation was conducted by Snowflake’s internal Coco team using dbt-bench, a specialized benchmark designed to measure an AI model's ability to autonomously author, debug, and maintain complex data transformation pipelines[1][3]. The rigorous testing suite consisted of 103 distinct data transformation tasks, with each model given three independent trials per task[1]. To succeed, the AI agents had to generate code that worked seamlessly across both DuckDB and Snowflake database architectures[2]. When looking at the overall success rate across three attempts, the performance difference was practically indistinguishable[1]. Zhipu AI’s GLM-5.2 achieved a Pass@3 score of 66 percent, landing just one percentage point behind Claude Opus 4.7’s score of 67 percent[1]. However, when the models were judged on their ability to solve a task correctly on the very first attempt, a clearer divergence appeared[1]. On first-attempt accuracy, Claude Opus 4.7 held a more commanding lead, scoring 53.7 percent compared to GLM-5.2’s 47.6 percent, indicating that Anthropic's model remains the more reliable and consistent option right out of the gate[1].

Despite achieving nearly identical cumulative success rates, the two models arrived at their solutions through dramatically different operational behaviors[1]. According to the Snowflake analysis, GLM-5.2 is significantly less efficient, requiring more interaction loops and executing a higher volume of tool calls to complete the same set of tasks[1]. On average, the Chinese model required 99 turns per task compared to 80 turns for Claude Opus 4.7[2][1]. In terms of tool execution, GLM-5.2 initiated an average of 40 execution-related tool calls per trial, whereas Opus 4.7 completed its trials with an average of 29[1]. This behavioral disparity had a massive impact on token consumption[1]. Across the entire benchmark run, GLM-5.2 consumed 860 million billing tokens, representing nearly double the 439 million tokens utilized by Claude Opus 4.7[2][1]. Snowflake researchers attributed this extreme token burn to a combination of more conversational turns, highly atomic API calls, and a significantly lower rate of prompt-cache reuse[1].

The benchmark also revealed stark differences in how each model manages validation and handles complex debugging loops[1]. A popular narrative in the AI community has suggested that GLM-5.2 owes its high benchmark scores to a more rigorous internal verification process. However, Snowflake’s findings paint a more nuanced picture, showing that the model validates differently rather than necessarily more meaningfully[1]. While GLM-5.2 excels at concurrently verifying code across multiple platforms, a capability Sridhar Ramaswamy identified as crucial for solving certain niche tasks that stumped other models[2], it is also highly prone to getting trapped in recursive, unproductive loops[2][1]. In one extreme example, GLM-5.2 executed 411 tool calls over a 24-minute period, obsessively checking database row counts, distributions, column types, and null values, only to ultimately fail all three attempts on the task[2]. In contrast, Claude Opus 4.7 resolved the exact same programming problem in just nine minutes using a highly concise 49 tool calls[2].

This efficiency deficit, however, is offset by the aggressive pricing structure of Zhipu AI's model, which completely upends the financial math of enterprise deployment. Anthropic currently prices Claude Opus 4.7 at $5.00 per million input tokens and $25.00 per million output tokens[4]. Zhipu AI offers GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens, which translates to roughly one-fifth the cost per output token of its Western rival[5][2]. Consequently, even though GLM-5.2 consumes nearly twice as many tokens per task due to its verbose, multi-turn approach, the overall financial cost of running the model remains dramatically lower than using Claude Opus 4.7[2][1]. For enterprises evaluating how to deploy autonomous coding agents at scale across thousands of database endpoints, these per-task economics are not merely a footnote. Instead, they represent a highly compelling cost-benefit Pareto frontier where slightly lower first-attempt consistency is heavily outweighed by an order-of-magnitude reduction in raw API costs[6].

The arrival of GLM-5.2 as a highly competitive, open-weight model carries profound implications for the global artificial intelligence sector. For years, leading Western AI labs like Anthropic and OpenAI have justified their soaring valuations by pointing to the exclusive capability moats of their proprietary, closed-source frontier models. However, the open-sourcing of GLM-5.2 under a permissive MIT license, complete with a stable 1-million-token context window and dual reasoning effort levels, demonstrates that international competitors are rapidly closing the gap[7][8][9]. Industry analysts suggest that this "peak close behind" phenomenon is eroding the pricing power of Western labs[10]. If open-weight models from international competitors can match closed-source performance on complex, long-horizon enterprise tasks at a fraction of the price, Western firms will face immense pressure to slash their profit margins or risk losing massive corporate contracts.

Ultimately, the findings shared by the Snowflake leadership signal a major paradigm shift in how corporations will architect their generative AI strategies moving forward. Instead of relying solely on expensive, single-model proprietary ecosystems, enterprises are increasingly positioned to leverage a highly competitive, multi-model landscape. This approach allows organizations to strategically deploy cheaper, highly capable open-weight models like GLM-5.2 for high-volume background tasks, while reserving ultra-premium proprietary models like Claude Opus 4.7 for tasks requiring maximum first-time precision[6][1]. As the capability gap between open-weight and proprietary models continues to shrink to a matter of mere months, the pricing pressure on Silicon Valley’s elite labs is only set to intensify, forcing a critical reckoning over the true commercial value of closed-source artificial intelligence.