Snowflake Adds Ability to Dynamically Route Prompts Across Multiple AI Models

Snowflake added dynamic AI model routing to its Cortex AI Gateway, enabling organizations to switch between proprietary and open-weight models, including China's DeepSeek-V4-Flash 0731 and GLM-5.3, while applying governance and security controls. In internal tests, AI agents using dynamic routing built a dbt pipeline with up to three times greater token efficiency than using a single frontier model, and engineering teams completed the same number of pull-requests with 25% greater token efficiency. Snowflake's AI Research Team found DeepSeek-V4-Flash scored 74.4% on data engineering tasks, outperforming a leading proprietary model, and GLM-5.2 scored 62.8% using fewer tokens than any other model tested.

TL;DR — Key Takeaways - Snowflake added dynamic AI model routing to Cortex AI Gateway, allowing organizations to direct prompts across proprietary and open-weight models. - The capability includes Chinese open-weight models such as DeepSeek-V4-Flash 0731 and GLM-5.3, alongside models from major Western providers. - Snowflake says dynamic routing can improve token efficiency while applying governance and security controls around model access. Snowflake this week revealed it has added a dynamic artificial intelligence AI model routing capability that enables organizations to switch between providers of proprietary and open-weight services. The Cortex AI Gateway that Snowflake makes available within its cloud service can now be used to dynamically route prompts between AI models https://www.snowflake.com/en/news/press-releases/snowflake-unlocks-better-ai-economics-dynamic-model-routing/ , including open-weight models such as DeepSeek-V4-Flash 0731 and GLM-5.3 from China, in a way that can also be used to limit the number of tokens consumed. Dynamic model routing has also been integrated across Snowflake’s flagship AI products, including Snowflake CoCo and Snowflake CoWork. Previously, Snowflake provided support for AI models from Anthropic, OpenAI, Google, xAI, Meta, and Mistral. As the cost of using more advanced proprietary AI models has increased, interest in using open-weight models as an alternative continues to rise, says Mayank Upadhyay, chief security and trust officer for Snowflake. While there are concerns that AI models from China may have been trained in a way that could create security issues, Snowflake has created a data governance framework along with additional controls, says Upadhyay. “We take care of security issues on behalf of the customer,” he adds. In internal Snowflake testing, AI agents using dynamic model routing with Cortex AI Gateway were able to build a dbt pipeline with up to three times greater token efficiency than using only a single frontier model. In a separate test, engineering teams completed the same number of pull-requests with 25% greater token efficiency. Snowflake’s AI Research Team also evaluated DeepSeek-V4-Flash on enterprise-focused tasks, with recent testing showing that DeepSeek v4 Flash outperformed a leading proprietary model by scoring 74.4% on data engineering tasks. GLM-5.2 also performed strongly at 62.8%, while using fewer tokens than any other model tested. Ultimately, organizations are going to want to be able to automatically mix and match workloads to the right model. Snowflake is making the case that organizations can use a platform where they already store much of their data to securely access AI models, regardless of how those models were built. It’s not clear how many organizations are relying on AI models developed in China, but a recent report from Stanford University https://techstrong.ai/features/stanford-report-us-losing-ground-to-china-in-ai-global-usage-growing/ suggests proprietary AI models are losing ground. At the same time, however, AI models from China have become a flashpoint as geopolitical tensions rise. The Trump administration is reportedly preparing to send a letter to allies informing them they should not be using AI models from China https://techstrong.ai/generative-ai/u-s-drafts-letter-warning-allies-to-pick-sides-in-ai-race-against-china-report/ if they signed on to a non-binding Pax Silica framework created last year. Each organization will need to determine for itself whether the potential reward for using AI models from China outweighs any potential risk, but if proprietary AI models continue to lose market share, the business models on which companies such as OpenAI and Anthropic operate may need to be revisited. One way or another, however, even in the age of AI it’s still the quality and security of the data that matters most rather than the underlying platforms used to process it.