Mark RadarMARK RADAR
About
中 EN
Sign in
Event File AI Large Language Models (LLMs) Open-Source AI Models

AT&T Cuts AI Costs 56% With Model Routing and Open Source

1 reports · First detected 2026-08-21 · Last active 2026-08-21

AT&T is reshaping its enterprise generative-AI stack around model routing, assigning employee requests to systems matched to their complexity and cost. Using LiteLLM as a common routing layer, the US telecommunications company can distribute workloads across proprietary and open-source models instead of relying on a single vendor. The approach offers a template for large companies seeking to contain rising AI inference expenses without materially weakening performance.

AT&T said the routing strategy has reduced costs for workloads including AI coding by as much as 56%, while performance declined only about 2%. Simpler requests are directed to cheaper or open-source systems, reserving proprietary models from providers such as OpenAI and Anthropic for more demanding tasks. The company plans to increase the share of usage handled by open-source models to between 60% and 70%, further limiting spending on commercial platforms.

All Coverage

1 original reports
PYMNTS.COM 2026-08-21
AT&T Slashes AI Costs by Adopting Model Routers and Open Source

The Backstory

The history behind this event
AT&T Cuts AI Computing Costs by 90% With Small Language Models2026-02-28 · 1 reports · similarity 0.86

AT&T shifted some tasks performed by its internal Ask AT&T assistant from large language models, or LLMs, to small language models, or SLMs. Selecting leaner models for specific business needs reduced the computing power and latency required for inference. The case shows that companies adopting generative AI do not always need to rely on large models to balance performance and cost.

AT&T said the switch to small language models cut Ask AT&T’s AI operating costs by 90%, tripled processing performance and improved system response times. However, reports did not disclose the actual amount saved, the model names, the formal deployment date or the testing period, meaning the results can currently be assessed only by their relative improvement.

Mark Radar｜MARK RADAR
Past Issues・by MarkReadFintech

If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →

All times are in Taipei time (GMT+8)