Snowflake claims 3x token efficiency from AI model routing

AI data platform company Snowflake claims its AI model routing technology can deliver up to three times greater token efficiency than using a single advanced AI model for workload, potentially helping enterprises contain the cost of deploying AI agents at scale.

The company said a test using its Cortex AI Gateway produced the efficiency gain while maintaining comparable output quality. The technology automatically directs tasks to suitable AI models based on their capabilities and the requirements of each workload, rather than relying on a single model for every request.

The capability addresses a challenge for businesses moving AI applications from pilot projects into production. AI agents can make multiple model calls and interact with tools to complete a task, adding to computing costs. Using a powerful model at every step can prove expensive when simpler models may be sufficient for routine work.

"The economics of AI are now boardroom math," Baris Gultekin, Vice President of Applied AI at Snowflake told Business Today. "As enterprises invest more in deploying AI across the business, they're starting to ask difficult questions about whether the value it creates justifies the cost," he said.

As AI usage scales across applications and agents, Snowflake argues that enterprises need to look beyond the cost of individual model calls and focus on what it calls "intelligence efficiency", or how effectively they turn compute, models, data and context into business impact. Small inefficiencies can compound across thousands of workflows, making the economics of model selection increasingly important as AI moves into production.

Snowflake announced dynamic model routing in August 2026 as businesses sought ways to make AI deployments more economical without compromising output quality.

Matching models to tasks

The principle behind dynamic model routing is to use a model suited to the task rather than defaulting to a more expensive option for every request. A routine summary, for instance, may not require the same reasoning capabilities as a complex analytical problem.

"The goal should be to make AI useful in more places across the business. That means being disciplined about where additional intelligence makes a difference and where it only adds cost," said Gultekin.

The economics become more complicated with AI agents, where completing a task can involve multiple model calls and interactions with tools. Enterprises need to assess not just the cost of each call, but also the quality of the final output, the time taken to generate it and the amount of correction required.

A cheaper model may not necessarily reduce the overall cost if its output requires substantial reworking. Model selection therefore involves balancing performance, latency and reliability against the expense of running the system.

Snowflake sees this model orchestration as part of a broader ‘agentic control plane’ for enterprises, designed to help manage models and agents while balancing cost, governance and performance at scale. Through dynamic model routing in Cortex AI Gateway, the system automatically selects the model suited to a task from models approved by the customer, based on considerations such as quality and cost.

The approach is intended to reduce the need for enterprises to use their most expensive model for every task, while maintaining the level of quality required for individual workloads. Gultekin says businesses should start by defining what a task requires and then evaluate models against those requirements.

"The obvious factors are quality, cost and latency, but others become important in production: reliability, availability, security, governance and how well the model works with the enterprise's own data and context," he said.

He cautions against relying solely on benchmarks to select models. Performance in a controlled test may not translate directly into business results because companies have their own data, operating rules and expectations. Enterprises need to evaluate models against actual workloads and revisit those choices as the technology evolves.

Governance remains critical 

Cost is only one part of the equation. As businesses use models from multiple providers, they also need to determine which models can access their information, where data can be processed and how requests are tracked.

Gultekin says governance must be built into model selection. Snowflake's routing system is designed to restrict model choices to those approved by administrators and provide visibility into which model handles each request, while respecting existing data security and residency controls.

The quality of the information supplied to a model also affects the outcome. Even a powerful model can produce inaccurate results when it lacks relevant business context.

"Giving an agent the right data, choosing an appropriate model for the task, and reducing unnecessary steps can all improve the outcome and its cost," said Gultekin.

Taken together, these factors form the basis of what Snowflake describes as intelligence efficiency: measuring AI not simply by model performance or token cost, but by how effectively compute, models, enterprise data and context translate into useful business outcomes.

Lower operating costs could also make it feasible for businesses to automate tasks that were previously too expensive to justify.

"Workflows that were too expensive to automate become practical. Teams can experiment more and bring AI into parts of the business that previously couldn't justify the investment," he said.