Snowflake targets cost of AI with dynamic model routing
With spending on AI spiraling out of control for many organizations, the vendor's latest new feature automates model selection to balance cost and performance.
Snowflake is doing something about the high cost of developing and managing agents and other AI applications.
Building an agent requires workflows such as model routing, context management, governing tool calls, batch data processing and network orchestration that can sometimes add up to well over $100,000, according to IT consulting firm Triple Minds. As a result, it is often significantly more expensive to develop AI tools than to create the traditional data products that have historically been the primary sources enterprises used to inform decisions and manage business processes.
With many organizations now spending far more to build AI tools than they did to develop data products, Snowflake on Tuesday unveiled dynamic model routing in its Cortex AI Gateway control layer for AI.
Dynamic model routing aims to reduce wasted spending on using bigger, more expensive models than are needed to inform many agents and instead automate model selection to strike the optimal balance between the quality of the model and the cost of the project.
"These capabilities are highly significant because they attack the operational overhead and suboptimal selection of models," William McKnight, president of McKnight Consulting, told TechTarget. "By automatically matching task complexity to the optimal model, dynamic routing eliminates the need for developers to rebuild their application infrastructure every time a new model is released."
Beyond dynamic model routing, Snowflake revealed that it plans to add models from DeepSeek and Z.ai to its Cortex AI development environment. The vendor already enables access to models from Anthropic, Google, Mistral, OpenAI and SpaceX. More model choices better enable users to strike the optimal balance between cost and performance.
Dynamic model routing and access to Z.ai's GLM-5.3 are not yet in private preview, while access to DeepSeek-V4-Flash is in private preview.
Customers want cost control
AWS in December made cost control the focal point of its data management initiatives given that AI development demands massive data workloads. Databricks and Microsoft also offer capabilities aimed at helping customers control their spending on AI.
Now, Snowflake is similarly addressing the cost of AI development and deployment.
As more enterprises move AI pilots into production, they're paying more attention to spending than they did when they were just trying to prove that something worked, according to Sanjeev Mohan, founder and principal analyst of analyst firm SanjMo. That is the reason data and AI providers are beginning to prioritize cost control.
Companies run AI across many apps and agents, and the default habit of sending every request to the most powerful frontier model turns out to be enormously wasteful, since most tasks don't need that horsepower.
Sanjeev MohanFounder and principal analyst, SanjMo
"Companies run AI across many apps and agents, and the default habit of sending every request to the most powerful frontier model turns out to be enormously wasteful, since most tasks don't need that horsepower," he told TechTarget.
For example, Canva was forced to cut its growth forecast from 30% to 20% because it relied too heavily on frontier models that proved more costly than expected, Mohan continued.
"The analyst read is that AI inference breaks the near-zero marginal cost that made software so profitable. … Managing spend across models is now worth billions," he said.
McKnight likewise suggested that as enterprises move more AI pilots into production, cost is becoming a greater concern, and vendors must respond to keep their customers.
"Enterprises are under pressure to transition speculative AI pilots into ROI-generating production and contain runaway cloud compute costs and the extreme engineering overhead of multi-vendor stacks," he said. "To prevent customers from abandoning these initiatives, database and orchestration vendors like Snowflake are prioritizing predictable pricing models that make AI development financially sustainable."
Customers, in fact, provided Snowflake with the impetus to develop dynamic model routing and expand the selection of models its users can choose from when building AI tools, according to Pavan Pothukuchi, the vendor's director of product for AI and machine learning.
He noted that with customers "becoming more rigorous about the economics" as AI moves from experimentation to production, Snowflake saw an opportunity to help them reduce spending by optimizing model selection.
"Customers weren't necessarily asking us specifically for dynamic model routing, but they were asking us to help improve the economics and performance of their AI deployments," Pothukuchi said. "Dynamic model routing is a natural response to that … so customers can focus on the outcomes they want from AI."
Addressing AI spending
Snowflake introduced Cortex AI Gateway in July 2025 to provide administrators with a single location to govern connections and interactions between agents, intelligently route workloads, and optimize AI use. Within Cortex AI Gateway, users can view token usage and costs, and establish spending limits for agents and other AI tools so they don't unexpectedly drive up costs.
The addition of dynamic model routing helps customers control their AI spending by automating the selection of the models that agents are built upon.
Snowflake noted that using the same model for every task can drive up costs since certain models are better suited for an agent's specific task than others. For example, one model may be better suited for informing customer service agents, while another may be better suited for supply chain optimization agents.
Dynamic model routing automatically funnels work that necessitates deep reasoning to frontier models while directing lower complexity tasks to less expensive models. In addition, routing decisions update as models and their pricing change, so users don't have to rebuild agents and other AI applications to lower spending.
"Dynamic model routing directly addresses the primary bottleneck of enterprise AI, which is the operational burden of manual model selection at scale and automatically matching each workload to the optimal AI model based on quality, speed, and cost," McKnight said.
Mohan likewise noted that the new feature directly addresses a substantial problem many enterprises face.
"The routing … removes a real operational chore, and does it automatically inside the platform where the data already lives," he said.
While dynamic model routing addresses customers that want to control costs, it does not distinguish Snowflake from its main competitors, according to Mohan. He noted that Microsoft offers Model Router in AI Foundry, Databricks provides routing through Unity AI Gateway, and vendors including OpenRouter and Vercel automate model routing.
However, its deep integration with Snowflake's broader data and AI platform, including tools such as CoCo and CoWork, along with governance capabilities such as role-based access, does provide differentiation.
"What's different about Snowflake's version is where it sits, routing inside Snowflake's governed boundary," Mohan said. "For a Snowflake-centric enterprise, routing that never moves data outside governance and attributes spending to the right team is the distinction, not the routing algorithm itself."
Meanwhile, Snowflake's overall cost of building and managing AI workloads is competitive, according to McKnight. Capabilities such as real-time concurrent LLM serving with a 0% error rate, optimized compute sizing, and Snowflake Container Services aid its price-performance ratio.
"Against the native services of major cloud providers, Snowflake is a highly competitive, well-balanced option for search," McKnight said.
With model routing among the reasons AI development and deployment costs are spiraling for some organizations, Snowflake introduced a feature that automates model selection to optimally balance cost and performance.
Looking ahead
Over the final months of 2026, Snowflake's plans include improving how the pieces that comprise the AI stack work together, according to Pothukuchi. Regarding dynamic model routing, he added that making the feature more adaptive will be a priority.
"We want to get increasingly sophisticated at understanding what a particular customer is trying to accomplish and selecting the intelligence best suited to that task, while continuously optimizing for quality and efficiency," Pothukuchi said.
Mohan suggested that Snowflake could build on dynamic model routing by adding cost control tools that go beyond routing to include generating and acting on their own recommendations.
"The natural next step is … moving from control to active guidance, which is exactly the muscle companies like Canva had to build by hand," he said.
McKnight, meanwhile, advised Snowflake to take steps to address the complexity of its platform and opaque billing that can catch customers off guard.
Specifically, it could turn its open-source pg_lake extension into a fully managed rival of Databricks' Lakebase, natively integrate an open semantic layer with Cortex AI and its Horizon Catalog to securely govern data mesh architectures, and add AI-driven FinOps controls along with micro-concurrency scaling to make spending more predictable.
"To continue holding its prime position in the market, Snowflake must focus on eliminating architectural complexity and billing anxiety," McKnight said.
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.
