AT&T Sends 40% of Employee AI to Open Models, Capping Anthropic Bill

Mark Austin, the vice president who oversees AI for AT&T's roughly 100,000 employees, says his internal platform now routes 40% of employee AI requests to open-source models instead of Anthropic's Claude or OpenAI's GPT. His target is 60% to 70% within a few years.

Ask AT&T, the internal assistant, processes about 45 billion tokens a day. A tool called LiteLLM sends the low-complexity work, which is most of it, to open-weight models like Nemotron, Llama and Gemma. Claude and GPT still handle the hard stuff. On coding workloads, the shift has cut costs by 56% with a 2% drop in quality.

"We've found open-source models to be on par with, and in some cases better than, older versions of the frontier closed models from Anthropic and OpenAI," Austin says. He also concedes the open models "trail frontier closed models by six to ten months," but says the gap is narrowing.

Ramp's corporate-card data points the same way. The share of AI-spending businesses paying model-serving platforms, the umbrella that also captures Chinese and open-weight models, rose from 4.5% in January to 6.1% by August. The median customer on those platforms spent about $248 per employee on AI in June, roughly 23 times the median AI-spending business. These are sophisticated buyers, not the budget end.

Ramp's other fresh number sits above all this: Anthropic edged past OpenAI in business spending share, 34.4% to 32.3%. The industry is telling on itself, too. Since late July, an "Open Weights and American AI Leadership" letter organized alongside Microsoft, Nvidia and Meta has collected more than 270 signatories, OpenAI among them.

Our Anthropic desk has been busy this week. Our open-source tracker has logged 275 stories in the last 90 days. This is the trend that could squeeze the company from below.