Snowflake announced on August 18 that it is adding dynamic model routing to Cortex AI Gateway. The feature inspects each request and sends simple work to lighter models, reserving frontier models for tasks that genuinely require deeper reasoning. In internal testing, the company reports up to 3x greater token efficiency at the same quality.
Using Expensive Models Only When They Actually Help
Once companies push AI agents into production, they tend to run everything through the same top-tier model. Routine summarization and complex design decisions get billed at the same rate. That is exactly the gap Snowflake set out to close.
Cortex AI Gateway, announced in July, serves as a unified foundation for governing agent connections, routing requests, and optimizing AI consumption. With dynamic model routing layered on top, the gateway itself selects the model with the best balance of quality and cost for each request. Development teams no longer need to hard-code a model name per call.
The reach extends beyond the gateway. Dynamic model routing is built into Snowflake CoCo and Snowflake CoWork, the company's flagship AI products, and is available to third-party AI agents that connect through Cortex AI Gateway.
Administrators keep meaningful control. They can decide which models and providers are exposed to their users, a design aimed at multinational organizations dealing with regional model availability and at regulated industries with strict compliance requirements. When model performance or pricing shifts, the gateway updates its routing decisions, so applications and agents do not have to be rebuilt.
3x on dbt Pipelines, 25 Percent on Pull Requests
Snowflake put two numbers behind the announcement, both from internal testing.
The first comes from an evaluation in which agents built a dbt data pipeline. Compared with a frontier-model-only path, dynamic model routing delivered up to 3x greater token efficiency while maintaining the same quality.
The second is closer to day-to-day software development: engineering teams completed the same number of pull requests with 25 percent greater token efficiency. Efficiency was measured using ADE-bench, a framework created by dbt for evaluating AI agents on real-world analytics and data engineering tasks.
Both figures come from the company's own tests, and Snowflake notes that results vary with workload and configuration. Still, the claim that mixing models cuts token consumption without sacrificing quality gives enterprises a reason to revisit their reliance on a single expensive model.
DeepSeek and GLM Join the Lineup, with Published Evaluation Scores
The model catalog is expanding as well. DeepSeek-V4-Flash 0731 and GLM-5.3 are being added to Snowflake Cortex AI, joining models from providers including Anthropic, OpenAI, Google, SpaceXAI, Meta, and Mistral.
What stands out is that Snowflake published evaluation numbers from its own AI Research Team. DeepSeek-V4-Flash scored 74.4 percent on data engineering tasks, outperforming the leading proprietary models used in the evaluation. GLM-5.2 also performed strongly at 62.8 percent while using fewer tokens than any other model tested.
Rather than treating each open model release as a one-off integration project, Snowflake says it is building an environment designed to keep pace with the open model ecosystem. Customers can adopt newer models without reworking their apps or infrastructure, and the same access controls and governance apply across open and proprietary options alike.
Availability comes in stages. DeepSeek-V4-Flash 0731 is already in private preview, dynamic model routing will enter private preview soon, and GLM-5.3 is also slated for private preview, though Snowflake notes this is subject to change based on model availability.
Tracking Who Spent What, Down to the Team
Automated routing is only half of the cost story. Cortex AI Gateway gives administrators visibility into token usage and costs, and lets organizations set spending limits across AI apps and agents.
Snowflake CoCo connects this to the existing role-based access control and tagging framework, so administrators can set default models, attribute usage to teams or cost centers, establish per-user quotas, and receive notifications as consumption approaches defined limits. For organizations that have lost track of whose budget AI spending comes from, this is the practical part.
Sridhar Ramaswamy, CEO of Snowflake, said enterprises are becoming far more rigorous about the economics of AI, and that the question has shifted from how much AI they are using to whether that AI is translating into meaningful business value. Snowflake calls this idea intelligence efficiency, a measure of how effectively a company turns compute, models, data, and context into business impact.
Sanjeev Mohan, Principal and Founder of SanjMo, made a related point: the real problem for enterprises is not the sheer number of model choices, but the operational overhead of picking the right one for every task at scale.
Summary
Snowflake has added dynamic model routing to Cortex AI Gateway, Snowflake CoCo, and Snowflake CoWork, giving the platform a way to match model choice to task complexity automatically. Internal testing reports up to 3x greater token efficiency on dbt pipeline construction and 25 percent on pull requests. DeepSeek-V4-Flash 0731 and GLM-5.3 join the model lineup, alongside usage visibility and per-team quota tools for cost management. It is the kind of update that starts to matter once agents move into production.
