US Companies Turn to Open Models as AI Bills Climb

US companies are starting to move some of their AI workloads off premium frontier models to cheaper ones, and some of that work is going to open-weight models. A Financial Times report published on September 27, 2026 names Tinder, PNC Financial Services, CH Robinson and Siemens among the businesses that have recently discussed using open-weight models. Three separate datasets from this month show the same trend, but they disagree on how much of it is going to open models specifically.

General Audience

What Companies Are Saying

The clearest example in the FT report is Tinder. Its chief technology officer, Vinay Kuruvila, said the company has started sending some queries from non-technical staff to open-weight models to control costs. “In January we were spending at the rate of $1 million per year and by July it had climbed to $10 million,” Kuruvila said, as quoted by PYMNTS. According to the FT, PNC, CH Robinson and Siemens have also talked about using open-weight models in recent weeks.

The companies are sending routine, high-volume queries to models that are good enough for the job, and keeping frontier models for work that needs them. This is happening as frontier API prices stay high. OpenAI’s GPT-6 Astra, released September 3, lists at $10 per million input tokens and $50 per million output tokens.

What the Data Shows

Survey data: open-weight usage is growing. A September 2026 survey of 200 respondents by Enterprise Technology Research, reported by Techstrong.ai, found that open-weight models now handle 34% of enterprise AI token usage, up from 23% a year earlier. Respondents expect that to reach 41% within 12 months. The share of respondents running open-weight models in production rose from 31% in July to 42%, and another 43% are running pilots. Cost savings was the most common reason, cited by 69% of production users. Unfinished security and compliance reviews were the biggest obstacle (62%), and only 10% said they were worried about model quality. Google’s Gemma was the most-used open model family in production (55%), ahead of Microsoft’s Phi (39%) and Meta’s Llama (35%).

Spending data: the shift is mostly to cheaper closed models. The September 2026 Ramp AI Index uses corporate card and bill-pay data from US businesses. It shows frontier models (Opus, Fable, Sol) falling from a 53% peak share of tokens in August to 45%. Ramp’s effective price per million tokens is down 41% from its March peak, to $0.68. But Ramp says most of the growth is going to cheaper closed models such as GPT-5.6 Terra and Anthropic’s Sonnet series. “We’ve heard from businesses who are imposing company-wide defaults that reduce usage of frontier models, saying standard models are still highly performant and also more cost effective,” wrote Ramp economist Ara Kharazian. Only 6.4% of AI-spending businesses in Ramp’s data pay for open-source models.

Capability data: open models are close behind. Mozilla’s State of Open Source AI v1.1, released in September, estimates that open models are 4.4 months behind closed ones in capability, using METR time-horizon data. It also finds they score about 3 points lower on benchmarks at roughly 60% of the price. According to the report, open-model capability doubles every 3.9 months, compared with 5.5 months for closed models. Closed models still lead in professional knowledge work, 1M-token long-context retrieval, and packaged compliance guarantees such as SOC 2 and HIPAA. The report also finds that open models reach production 12 percentage points less often than closed ones, and it attributes that gap to tooling rather than capability.

What This Means

Put together, the three datasets suggest companies are doing more cost-based routing, and open-weight models are only part of that. Most of the savings so far seem to come from moving to cheaper models from the same closed vendors. That change takes one line of configuration. Moving to open weights also means hosting the model, running security reviews and maintaining it over time. ETR’s respondents rank those costs well above model quality as a concern. Ramp’s data may also undercount open-model use, because self-hosted models appear as cloud compute bills rather than AI subscriptions. That could help explain why the survey and spending figures differ.

For universities and research groups, the lesson is similar. The model that scores highest on a benchmark is not always the right choice. For many teaching, research-support and administrative tasks, a smaller closed model or an open model running on the institution’s own hardware may do the job at a much lower cost. Running the model in-house also keeps the data under the institution’s control.

Related Coverage

• OpenAI Releases GPT-6 Sol and Luna at Half the Price: the closed mid-tier models that compete for the same workloads
• Mozilla Launches Thunderbolt: An Open-Source, Self-Hostable Enterprise AI Client: tooling for companies that want to self-host
• Qwen3.8-2.4T-A95B: Alibaba Open-Weights Its Max-Tier Flagship: the largest open-weight language model released so far

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

• Financial Times: US corporate shift to open-weight AI models (September 27, 2026)
• PYMNTS: Businesses Embrace Open-Weight AI Amid Heavy Tech Costs
• Ramp: September 2026 Ramp AI Index
• Techstrong.ai: Open-Weight AI Models Expected to Capture 41% of Enterprise Token Usage
• Mozilla: The State of Open Source AI, v1.1 (September 2026)