AT&T Reduces AI Costs by 56% Through Model Routers

AT&T has reduced the expenses associated with coding and other advanced artificial intelligence tasks by as much as 56 per cent. The telecommunications company achieved these savings by implementing tools that direct employee queries to less expensive models when the task complexity allows. According to a report from The Information published on 20 August, this strategic shift resulted in a negligible decline in performance quality, with output decreasing by only 2 per cent.

The cost reductions were facilitated by the adoption of LiteLLM model routers. These tools assess the complexity of each request and determine whether it can be processed by a cheaper AI model. Mark Austin, a vice president at AT&T who oversees the company’s internal AI usage, confirmed these figures in an interview cited by the report. The company is actively working to stabilise its spending on models from Anthropic and OpenAI. To achieve this, AT&T is increasing its reliance on open-source or open-weight models. The firm intends to raise the proportion of employee queries handled by these open-source solutions from the current 40 per cent to between 60 per cent and 70 per cent in the coming years.

AT&T is currently utilising open-source models such as Nvidia’s Nemotron, Meta’s Llama and Google’s Gemma. While the company is evaluating the potential risks associated with models from Chinese firms DeepSeek and Moonshot, it is not currently using them. Austin noted that open-source capabilities have historically lagged behind frontier models by six to 10 months. However, he stated that this gap is narrowing, with open-source options now performing as well as or better than older models from Anthropic and OpenAI. This move comes as businesses seek to manage rising AI costs, which have been driven by the shift from chatbots to agents and the transition to token-based billing.