Cheaper tokens, bigger bills. Why companies are moving to open-weight models

Imagine your AI bill goes from $1 million in January to $10 million in July. That is what happened to Tinder's AI spending this year, according to the Financial Times. It is an extreme example. But it shows a problem that I now see in many organisations.

At the weekend, the FT reported that corporate America is turning to lower-cost open-weight models to control IT costs. The shift is most visible in technology firms, but PNC Financial Services, CH Robinson and Siemens have also discussed open-weight models in recent weeks. Mentions of open-weight or open-source models on US earnings calls rose sixfold year on year in August and September, per AlphaSense data.

Frontier models getting better and cheaper

This is strange at first. The frontier models are getting better, and they are also getting cheaper. Anthropic released Claude Opus 5.5 on 22 September at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Anthropic says the model costs about 40% less on typical workloads, because it also uses fewer tokens per task. In the same week, OpenAI released GPT-6 Sol and GPT-6 Luna at prices at least 50% below GPT-5.6. Epoch AI estimates that the price of a constant level of AI performance falls by approximately 13 times each year.

So why are companies looking for alternatives? Because usage is growing faster than prices fall, their bills are still going up.

Open source and security

I expected large enterprises to be slow to adopt open-weight models. Security teams, procurement and legal all have questions. The data says the opposite. On Vercel's AI Gateway, open-weight models ran 56% of all tokens in August, up from 7% in December. OpenRouter shows the same pattern. Chinese models took between 57% and 67% of OpenRouter tokens in the week of 14 September, up from 6% to 13% in February. AT&T now runs 40% of its AI workloads on open models and wants to reach 70% within a year.

Security concerns are valid however. Two House committees opened a joint investigation in April 2026, with letters to Cursor and Airbnb, and DoorDash received a similar inquiry by July. But no federal law or executive order stops private US companies from using Chinese models. And the risk depends on how you deploy these models. A model called through a hosted Chinese API has very different exposure from a self-hosted copy of the same model. OpenRouter now offers US in-region routing, so requests are processed only on US-based providers.

Open source adoption

For me, the most important number is this one. Open-weight models handled 56% of Vercel's tokens in August but took only 14 cents of every dollar spent, while Anthropic took 64 cents. The companies that do this well are not leaving the frontier models. They get more inference from the same budget and keep the frontier models for tasks that justify the higher price.

That is a finance decision as much as a technology decision. The question for a CFO is not "which model is best?" It is "which model does each task need, and what does each task cost?" Most organisations cannot answer that today. That is the problem we built dataintelligence to solve.