Snowflake adds dynamic model routing to cut AI inference costs

Snowflake adds dynamic model routing to cut AI inference costs

Snowflake announced dynamic model routing capabilities within its Cortex AI Gateway designed to reduce enterprise AI spending by automatically selecting the most cost-effective model for each task. The company also plans to add DeepSeek-V4-Flash 0731 and GLM-5.3 to its model portfolio, according to details first reported by Snowflake.

The new routing system directs simpler or repetitive tasks to efficient models while reserving frontier models for work requiring deeper reasoning. Snowflake says the approach addresses a growing challenge as companies deploy more AI applications: using the same expensive model for every task drives up costs, while manually evaluating and managing multiple models creates operational overhead.

Dynamic model routing is integrated across Snowflake's AI products including CoCo and CoWork, and is available to third-party agents using Cortex AI Gateway. The system gives customers control over which models and providers are available to users—particularly important for organizations navigating regional availability or compliance requirements.

Performance and efficiency gains

Snowflake's internal testing showed agents using dynamic model routing with Cortex AI Gateway built a dbt pipeline with up to three times greater token efficiency compared to using only frontier models, while maintaining equivalent quality. In a separate test, engineering teams completed the same number of pull requests with 25 percent greater token efficiency.

The company's AI research team evaluated DeepSeek-V4-Flash on enterprise tasks, with recent testing showing it scored 74.4 percent on data engineering tasks—outperforming the leading proprietary models used in the evaluation. GLM-5.2 scored 62.8 percent while using fewer tokens than any other model tested, based on ADE-bench, a framework created by dbt for evaluating AI agents.

Expanded model access and cost controls

Snowflake is adding DeepSeek-V4-Flash 0731 and GLM-5.3 to its existing portfolio, which includes models from Anthropic, OpenAI, Google, SpaceXAI, Meta, and Mistral. Both new models are expected to enter private preview, though GLM-5.3 availability is subject to change.

Cortex AI Gateway now provides administrators visibility into token usage and costs, with tools to establish spending limits across AI applications and agents. Snowflake CoCo extends these controls through role-based access, allowing administrators to set default models, attribute usage to teams or cost centers, establish per-user quotas, and receive notifications as consumption approaches defined limits.

Why it matters

As enterprises scale AI deployments, inference costs can spiral when organizations default to expensive frontier models for all tasks. Snowflake's automated routing addresses a practical bottleneck: development teams lack the time and expertise to manually optimize model selection for every request. By absorbing this complexity at the platform level, Snowflake enables organizations to improve what it calls "intelligence efficiency"—the ratio of business value to AI compute spending—without rebuilding applications as the model landscape evolves. The addition of high-performing open models like DeepSeek-V4-Flash provides cost-effective alternatives that internal testing suggests can match proprietary model quality on specific enterprise workloads.

Snowflake CEO Sridhar Ramaswamy said enterprises are becoming more rigorous about AI economics, focusing on whether AI translates into meaningful business value rather than simply measuring usage volume.

The dynamic model routing feature was announced August 18, 2026, and is entering private preview. Details were first reported by Snowflake in a press release.