Snowflake Adds Dynamic Model Routing to Cut Enterprise AI Costs

Snowflake is adding dynamic model routing to its Cortex AI Gateway, allowing enterprises to automatically match AI workloads with models based on quality, speed, cost and customer preferences.

The company announced the capability on August 19, alongside plans to add DeepSeek-V4-Flash 0731 and GLM-5.3 to Snowflake Cortex AI. The updates are aimed at giving enterprises more control over AI inference costs as they deploy more models and agents.

“Enterprises are becoming much more rigorous about the economics of AI. The question is no longer how much AI they are using, but whether that AI is translating into meaningful business value,” said Snowflake CEO Sridhar Ramaswamy.

Dynamic model routing is available through Cortex AI Gateway and is integrated into Snowflake CoCo and Snowflake CoWork. It can also be used by third-party AI agents connected to the gateway.

Under the system, lower-complexity or repetitive requests can be sent to more efficient models, while tasks requiring deeper reasoning can be routed to frontier models. Customers can also control which models and providers are available to users, which Snowflake says can help organisations manage regional availability and compliance requirements.

“Achieving intelligence efficiency requires the flexibility to use the best model for each task as the landscape evolves,” Ramaswamy said. “Snowflake’s role is to absorb that complexity so customers can focus on outcomes while we optimise model choice underneath.”

Snowflake said internal testing found that dynamic routing could improve token efficiency without reducing output quality. In one evaluation, agents building a dbt pipeline achieved up to three times greater token efficiency compared with using a frontier model for every task. 

In another test, engineering teams completed the same number of pull requests with 25% greater token efficiency.

The company is also expanding the model options available through Cortex AI. DeepSeek-V4-Flash 0731 and GLM-5.3 will join models from Anthropic, OpenAI, Google, SpaceXAI, Meta and Mistral.

Snowflake said its AI Research Team recently evaluated DeepSeek-V4-Flash on enterprise-focused tasks, where the model scored 74.4% on data engineering tasks. GLM-5.2 scored 62.8% while using fewer tokens than the other models tested.

The company is also adding controls for managing AI consumption. Cortex AI Gateway gives administrators visibility into token usage and costs, while Snowflake CoCo allows organisations to set default models, assign usage to teams or cost centres, establish per-user quotas and receive alerts when consumption approaches defined limits.

Sanjeev Mohan, principal and Founder of SanjMo, said the challenge for enterprises is increasingly the operational work involved in selecting models for different workloads.

“Snowflake’s dynamic model routing directly addresses that gap,” Mohan said. “The ability to match workload complexity to model cost, without rebuilding your infrastructure every time a new model drops, is exactly the kind of efficiency enterprises need to move from AI experimentation to AI at scale.”

The move comes as enterprises face a growing number of proprietary and open models, with differences in performance, pricing and availability. Snowflake is positioning Cortex AI Gateway as a layer that can manage those choices while keeping enterprise data and access controls within its platform.