Snowflake introduced dynamic model routing to cut AI costs by matching tasks to cheaper models.
It also added open models DeepSeek-V4-Flash and GLM-5.3.
Internal tests show up to 25% fewer tokens used.
What happened
Snowflake announced dynamic model routing in Cortex AI Gateway, which sends each request to the most affordable model that can confidently complete it. The company also expanded its open model portfolio with DeepSeek-V4-Flash 0731 (in private preview) and GLM-5.3 (coming soon).
Why it matters
In internal tests, dynamic routing handled a dbt pipeline workload with up to three times greater token efficiency than a frontier-model-only approach, and a coding test kept the same pull-request throughput while using about 25% fewer tokens. On ADE-bench, DeepSeek-V4-Flash scored 74.4%, beating the leading proprietary model tested.
What to watch
Snowflake serves these open models itself, keeping inference inside its secure perimeter. GLM-5.3 is self-hostable and designed to keep data in the customer's environment, per the announcement.
Ask the AI about this article →
The announcement ties two trends together: model selection and cost control. Snowflake argues the key question is not which model is best, but which is best suited to each task. Dynamic routing operationalizes that by choosing the cheapest adequate model per request, reducing reliance on expensive frontier models for routine work.
The open model additions matter because a router's value depends on the pool of options. DeepSeek-V4-Flash's 74.4% score on ADE-bench, outperforming the leading proprietary model tested, supports the claim that open models are competitive where data teams work. GLM-5.2's past scores—66% on ADE-bench with the lowest token footprint—suggest a similar trade-off. Snowflake serving these models itself, rather than proxying third-party APIs, aims to keep inference within its governance boundary.
Snowflake frames the long-term benefit as compounding gains in what it calls intelligence efficiency. Routing decisions become more precise as models are evaluated across workloads, and a growing pool of efficient models shifts more work to cheaper options. The system is designed to lower the average cost per outcome over time without requiring customers to rebuild applications or manually manage routing logic.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
