Corporate America Quietly Routes AI Work to Cheaper Open-Weight Models as API Bills Soar

Why this matters: the $10-million API bill is forcing companies to treat frontier models as a luxury, not the default.

**September 28, 2026** — American companies are quietly shifting their everyday AI workloads away from expensive frontier models toward cheaper **open-weight models**, according to a **Financial Times** report published Sunday that was summarized by **PYMNTS** on September 27. The shift, driven by runaway API costs, marks the first broad corporate pushback against the assumption that the most powerful models must handle every task.

Image: Data center server racks glowing blue, symbolizing the AI infrastructure behind the open-weight cost shift. Credit: Wil Weterings / Wikimedia Commons, Public domain.

**Editor's Highlight:** The smartest model is no longer the default model: procurement teams are routing routine work to whatever is cheapest and good enough, and saving the frontier giants for the few tasks that truly need them.

Table of Contents

• Key Facts
• The Tinder bill shock
• AT&T's router: 40% of requests now go to open models
• Open-weight models are catching up
• What this means for OpenAI and Anthropic
• Bottom Line
• FAQ
• Sources

Key Facts

Detail	Information
Trend	Companies routing routine AI work to cheaper open-weight models
Reported by	Financial Times, September 27, 2026 (via PYMNTS)
"Open weight" mentions	Up sixfold in Aug-Sep earnings calls vs. 2025 (AlphaSense)
Tinder's AI spend	$1M/yr run rate in January to $10M by July
AT&T open-model share	40% of employee AI requests, targeting 60-70%
AT&T cost savings	Up to 56% on coding tasks, ~2% quality dip
Open-weight examples	Meta Llama, Alibaba Qwen, Z.ai GLM, Google Gemma, Nvidia Nemotron

The Tinder bill shock

**Vinay Kuruvila**, chief technology officer at **Tinder**, told the Financial Times the dating app has started routing some queries from non-technical users to open-weight models in an effort to get a handle on rising AI costs. The numbers explain why.

"In January we were spending at the rate of **$1 million per year** and by July it had climbed to **$10 million**," Kuruvila said. "I don't want another 10X increase."

His conclusion cuts to the heart of the shift: "The frontier models like **OpenAI's Astra** and **Claude Fable** already have enough intelligence for **90% of the tasks** we're trying to do. If open-weights models catch up, I may not need to use them anymore."

Tinder is not alone. The FT found businesses across sectors, including **PNC Financial Services**, logistics company **CH Robinson** and industrial giant **Siemens**, discussing open-weight models in recent weeks. Management references to "open weight" or "open source" models in earnings calls and investor conferences **jumped sixfold** in August and September compared with the same two-month stretch in 2025, according to data from research platform **AlphaSense**.

**Newsroom Focus:** This is procurement logic, not ideology. As PYMNTS put it, for a CFO the contest is "less ideological than financial": the question is whether savings, flexibility and control justify running more infrastructure yourself.

AT&T's router: 40% of requests now go to open models

The clearest example of the new playbook comes from **AT&T**. The telecom giant built its internal AI platform, **Ask AT&T**, on top of the open-source routing tool **LiteLLM**, which scores every incoming request for difficulty before deciding where to send it.

Simple queries go to open-weight models such as **Meta's Llama**, **Google's Gemma** and **Nvidia's Nemotron**. Anything where a wrong answer is expensive, like complex code generation, still goes to a frontier model.

**Mark Austin**, the AT&T vice president who oversees AI for the company's roughly **100,000 employees**, said open models now handle **40% of employee AI requests**, with a target of **60% to 70%** within a few years. On coding tasks specifically, the routing cut costs by as much as **56%**, with only about a **2% dip in quality**. Ask AT&T now processes roughly **45 billion tokens a day**, according to Startup Fortune's reporting on the FT analysis.

According to data from **SiliconData** cited by the FT, inference token spend has been moving away from closed models, the ones OpenAI and Anthropic sell through their APIs, since **mid-July**. The pattern is not companies canceling frontier contracts outright, but quietly sending the bulk of everyday requests somewhere cheaper.

Open-weight models are catching up

The economics only work because the capability gap is narrowing. Open-weight models, which users can download and run on their own hardware, have been closing in on closed models for months, and token costs across the board keep falling.

**Analysis Note:** "Open-weight" is not the same as fully open source. The weights are published and the model can be run locally, but training data and methods often stay closed. For a CFO, what matters is the price and the control, not the philosophy.

The shift also has a geopolitical flavor: reports this year have noted companies routing work to models from Chinese firms such as **Alibaba's Qwen** and **Z.ai's GLM**, alongside Meta's Llama, as the cheapest "good enough" option.

What this means for OpenAI and Anthropic

The trend showed up in **Ramp's September 9 AI Index**. Based on August 2026 spending data, **43.8%** of US businesses in Ramp's sample purchased Anthropic products, compared with **39.8%** for OpenAI. Ramp pointed to the growing use of cheaper models as a direct threat to AI companies that rely on businesses steadily spending more on frontier systems.

The findings line up with the **OECD's agentic AI report** published in September 2026: interviews with 25 organizations found cost savings were a common reason for using agentic systems, alongside higher productivity and help with labor shortages.

**Practical Takeaway:** For enterprise AI buyers, the emerging best practice is tiered: define which tasks genuinely need frontier reasoning, and route everything else to the cheapest model that clears the bar. For the labs, the pressure is now on pricing, not just benchmarks.

Bottom Line

The open-weight shift is the market's answer to the AI bill shock: stop renting the smartest model for every query and start routing like it costs money, because it does.

What to watch: whether OpenAI and Anthropic cut prices or add cheaper tiers to defend enterprise share, and whether open-weight models keep closing the capability gap fast enough for Tinder's CTO to make good on his threat.

FAQ

What are open-weight AI models?

Open-weight models publish their trained parameters (weights) so anyone can download and run them on their own hardware. Examples include Meta's Llama, Alibaba's Qwen, Google's Gemma and Nvidia's Nemotron. They are typically far cheaper to run than API access to closed frontier models.

Why are companies switching to them now?

API costs have exploded as AI usage scales: Tinder's spend went from a $1 million annual run rate in January to $10 million by July. Open-weight models handle routine tasks well enough at a fraction of the price.

How big is the shift?

Management mentions of "open weight" or "open source" models in earnings calls jumped sixfold in August-September versus 2025, per AlphaSense. AT&T says open models now handle 40% of its employee AI requests, targeting 60-70%.

Does quality suffer?

AT&T reported only about a 2% quality dip on coding tasks after routing to open models cut costs by up to 56%. The strategy keeps frontier models for high-stakes work where errors are costly.

What does this mean for OpenAI and Anthropic?

Ramp's September AI Index flagged growing cheaper-model use as a threat to labs that depend on ever-rising enterprise spending. Expect pricing pressure and possibly cheaper tiers as the labs defend their customer base.

Sources

• PYMNTS, "Businesses Embrace Open-Weight AI Amid Heavy Tech Costs," September 27, 2026 — FT report summary, Tinder CTO quotes, AlphaSense data: https://www.pymnts.com/news/artificial-intelligence/2026/businesses-embrace-open-weight-ai-amid-heavy-tech-costs/
• Startup Fortune, "Companies are quietly routing AI work away from OpenAI and Anthropic to cheaper models" — AT&T Ask AT&T details, SiliconData inference-spend data: https://startupfortune.com/companies-are-quietly-routing-ai-work-away-from-openai-and-anthropic-to-cheaper-models/
• CryptoCompass, "Cheaper AI models gain ground as corporate America cuts costs" — Ramp AI Index figures, OECD agentic AI report: https://cryptocompass.com/articles/cheaper-ai-models-gain-ground-as-corporate-america-cuts-costs
• GetLatka, "How This $3.6M/yr Startup Runs Open AI Models" (YouTube) — on-topic explainer embedded above: https://www.youtube.com/watch?v=MaqmkzQaWso
• Wikimedia Commons, "Datacenter" by Wil Weterings (public domain) — hero image: https://commons.wikimedia.org/wiki/File:Datacenter.jpg