On September 27, PYMNTS published a piece on how American companies are rethinking their AI model choices as costs climb — built around a candid admission from Tinder CTO Vinay Kuruvila: "In January we were spending at the rate of $1 million per year and by July it had climbed to $10 million… I don't want another 10X increase." The piece echoes Vercel's September AI Gateway report: in August, open-weight models overtook closed proprietary models in token volume for the first time — 56% versus 44% — up from just 7% back in December 2025.
There's an important detail easy to miss here: despite that, open models captured only 14% of total business AI spend — the bulk of the money (86%) still goes to top-tier proprietary models like Claude or GPT. In other words, this isn't a flight from quality models; it's a division of labor. The expensive, top-tier model stays for tasks that demand maximum accuracy, while routine, high-volume operations — classifying requests, extracting data from documents, drafting text — are migrating en masse to cheaper open models. The average token price dropped 23.2% in August, the steepest fall since April.
What this means in terms of an actual automation budget
Picture a mid-sized logistics company (PYMNTS cites CH Robinson in a similar situation) that processes incoming freight requests via AI: parsing emails and PDFs for shipment details, checking the route, drafting a preliminary quote. If a year ago the whole chain ran on a single expensive "do-everything" model, a sensible architecture today is a multi-model pipeline: a cheap open model pulls data out of the email and PDF, while a pricier model checks the final quote before it goes out to the client. Rough estimates put the savings from a setup like this at 40-60% off the monthly AI bill, with no quality loss at the critical last step — which is exactly the kind of effect high-volume customers describe in the PYMNTS piece.
Where the savings stop and a new headache begins
There's a less rosy side too, one that tends to get left out of the more breathless retellings of the report. Open models almost always mean the company takes on infrastructure responsibility itself: where to host the model, who tracks security updates, who guarantees uptime under peak load — instead of simply paying for an API and washing your hands of those questions. For a business without its own engineering team, that can turn into hidden costs that don't show up in a simple price-per-token comparison. The article is blunt about it: Tinder's CTO admits that current top-tier models already exceed what most of the company's tasks actually require — but adds the caveat "if open models keep progressing," meaning the decision is still a bet on the future, not a guaranteed saving today.
For a small or medium business owner in Portugal, the practical takeaway isn't to rush out and switch to open models on your own — it's that the cost of well-designed automation keeps falling, and a year-old verdict of "too expensive for our volume" is worth recalculating. The difference in architecture — one model for everything versus a multi-model pipeline — often shapes the final bill more than the choice of vendor does.
There's also a telling dynamic inside the proprietary-model market itself, which Vercel's data captures: after Claude Opus 5 launched, its predecessor Fable 5 fell from 13.2% to 4.9% of spend share in a single month, and nine in ten teams previously running Fable cut their usage almost immediately once a cheaper alternative at the same tier appeared. That shows price competition is just as fierce inside the "expensive" tier as it is on the open-model side — and companies that revisit their AI stack quarterly, rather than annually, end up in a noticeably better spot on the final bill.
At Dayava, we design exactly these kinds of multi-stage pipelines around each client's actual budget, rather than selling one universal solution. If a similar solution interests you for your business, leave a request at dayava.pt/contactos/.
Source: PYMNTS, "Businesses Embrace Open-Weight AI Amid Heavy Tech Costs"