Coinbase, Snowflake Turn to Model Routers to Cut AI Bills
TL;DR
- Coinbase reportedly cut internal AI spending by roughly half while token usage kept growing, mostly by routing prompts to cheaper open-weight models.
- Coinbase pushed its caching hit rate from about 5 percent to 60 percent, and 91 percent of engineers had never hit the old usage limits.
- Snowflake and Palo Alto Networks told The Information that swapping in cheaper models on specific tasks produced considerable savings.
The interesting shift in enterprise AI this quarter is not another frontier release, it is that finance teams have started asking a very awkward question about the frontier bill. In reporting from The Information, the answer that keeps coming back is model routing, sending each prompt to the cheapest model that can actually handle it rather than defaulting to the most capable one.
The piece sorts what is on the market into five categories: OpenAI's built-in switcher inside ChatGPT, independent products like OpenRouter (whose auto mode is powered by the startup Not Diamond), cloud-provider modules such as Databricks' Unity AI Gateway, custom stacks built by enterprise IT teams, and DIY setups engineers assemble themselves. The pitch is boring on purpose: an email summary does not need a frontier reasoner, and paying frontier rates for it adds up fast.
The receipts are the interesting part. Coinbase, according to coverage summarizing the same shift, cut internal AI spending by nearly half while token usage kept growing, mostly by routing prompts through cheaper open-weight models and pushing its caching hit rate from about 5 percent to 60 percent. It did that without imposing usage caps, and CEO Brian Armstrong noted that 91 percent of engineers had never hit the old limits anyway. Snowflake and Palo Alto Networks separately confirmed to The Information that using cheaper models for specific tasks led to considerable savings, and Databricks CEO Ali Ghodsi told the outlet the Unity AI Gateway is popular because customers are 'exhausting their budgets at too rapid a pace.'
The honest caveat is that a router is only as good as its evals, and every one of these deployments is somebody's private benchmark rather than a controlled test. The reporting does not give you failure modes, quality regression rates, or what happens when a cheap open-weight model quietly gets a hard question wrong in production. Take the specifics as reported, not settled. What is clearly shifting is the buying posture. Instead of one frontier contract per company, procurement is starting to look like a portfolio, and the winners over the next year will probably be whichever vendor makes the routing layer feel like plumbing rather than a science project.
Originally reported by theinformation.com
Read the original article →Original headline: The Five Kinds of Model Routers That Cut AI Costs