Tech companies move to open AI models
The Pragmatic Engineer
Tech firms are swapping pricey AI models for open ones. Uber, Pinterest and AT&T say the switch cut bills fast, sometimes with little quality loss.
Based on reporting by The Pragmatic Engineer — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Tech companies are starting to treat AI spend like any other budget line: something to squeeze, not worship. A few months ago, the playbook was already taking shape — cheaper open models, smarter routing, stricter usage limits, and more talk about not burning frontier-model money on simple tasks. Now several big names say those tactics are working.
Uber is the clearest example in the source. The company had already blown through its annual AI budget in the first three months of the year, and then engineering teams went after the problem hard. Uber says it cut cost per AI request by 34% and cost per AI session by 52%, while overall usage kept rising. Since March, spend has stayed flat even as token use and session count climbed.
The biggest lever looks to have been open weight models. Uber says those models can be 2-20x cheaper than frontier ones, and on code review the gap is stark: the priciest open weight model in Uber’s example cost $0.30 per review, versus $0.50 for the cheapest frontier model and $2.50 for the most expensive. But the savings came from a pile of smaller changes too: weekly benchmarking on real work, better model choice, cheaper subagents for simpler jobs, medium-effort defaults, automatic compaction above 400K tokens, and prompt caching in Uber’s own harness, Minions.
Pinterest and AT&T tell a similar story. Pinterest’s CEO said open source models inside its own secure cloud infrastructure are delivering better results for Pinterest-specific use cases, and that cost per transaction is now under 8% of comparable closed proprietary models. AT&T says it cut its AI bill by 56% while seeing only a 2% drop in output quality after moving workloads to open models, with developers still using stronger models for harder work and cheaper ones for simpler tasks.
The broader pattern is hard to miss. Anthropic’s models, especially Opus 5, are now far pricier than some alternatives, and that seems to be pushing companies toward open weight providers on inference. Databricks’ summary of experience at Stripe, Coinbase, Uber and Ramp points the same way: open models bring the biggest savings, then smarter routing, then the more boring controls. Boring is winning.
My take — AI-written commentary, not fact-checked reporting
The AI bill is finally meeting the real world, and the real world is rude about it. Companies don’t need a spiritual relationship with frontier models; they need work done cheaply, and open models are getting the call because math still works better than marketing. Somehow that remains a surprise in 2026, which says more about the hype cycle than the tech.
Read more about this at: The Pragmatic Engineer