Six large firms cut AI bills by routing work to open models

Large technology companies have started moving simpler AI workloads off proprietary models and onto open ones.

Uber, Pinterest, Stripe, Coinbase, Ramp and AT&T are all named as doing it, in reporting by Gergely Orosz's newsletter The Pragmatic Engineer.

The reported saving is roughly 50% on AI bills. That figure is the newsletter's, drawn from the cost-cutting work at those companies.

The mechanism it describes has two parts. The companies dropped proprietary models where they could, and used model routing to decide which request goes where.

The logic is easy enough to state. Not every request needs the most expensive model available, and paying frontier prices for classification, extraction or summarisation is a choice rather than a requirement.

What the newsletter does not set out publicly is the triage rule itself: which workloads each company judged simple enough to move, and at what point output quality stopped being acceptable. That part sits behind its paywall.

No per-company figures have been published either. The 50% is a headline number across the trend, not a line item for any one of the six.

For a founder, the useful signal is that the companies doing this are not small. Uber, Stripe and AT&T run AI at volumes where a routing layer earns its engineering time quickly; a seed-stage company with a four-figure monthly bill may not.

The work is also not free. Routing means running an evaluation set, deciding what a good answer looks like for each task, and monitoring the cheaper model for the day it starts getting things wrong.

The cheapest version of this experiment is to take the highest-volume, lowest-stakes prompt in your product, run it against an open model, and compare the outputs against the ones you are paying for now.

If the difference is invisible to your users, you have found where your bill was optional.