# How AT&T cut AI coding costs 56% with open-weight models and LiteLLM > Source: > Published: 2026-08-28 19:22:12+00:00 ## The numbers AT&T cut the cost of coding and some other advanced AI tasks by **as much as 56%, with quality declining only 2%**, according to The Information (August 20, 2026), citing an interview with Mark Austin, the AT&T VP who oversees AI used by the company’s employees. The mechanism was not a new model - it was two boring, replicable moves: a **model router** (LiteLLM, open source) that judges task complexity and sends the easy queries to cheap models, and a deliberate shift of volume onto **open-weight models**. The routing part matters as much as the models. Not every query needs a frontier model - most do not. LiteLLM sits between the employee’s tool and the model providers, assesses the task, and routes accordingly. The 56% saving is what happens when the expensive model only answers the queries that actually need it. ## The open-weight share is climbing Today **40% of AT&T employee queries run on open-source or open-weight models**, with a target of **60-70% in the coming years**. The models named: Nvidia’s **Nemotron**, Meta’s **Llama**, and Google’s **Gemma**. Spending on Anthropic and OpenAI is targeted to stay flat - growth gets absorbed by open weights, not bigger frontier bills. Two notable exclusions: AT&T is not using open models from Chinese labs - no DeepSeek, no Moonshot - while it evaluates the risks. And Austin’s quality read: open-source capabilities have historically lagged frontier models by **six to 10 months**, the gap is narrowing, and today’s open models are “just as good or better” than older Anthropic and OpenAI models. ## The enterprise stamp on an argument we keep making This is the first large-enterprise number we have seen that quantifies what the independent evals have been suggesting all year: the open-weight tier is no longer a compromise play, it is a purchasing strategy. It lines up with what we measure on this site - [Nemotron 3.5 Lightning](/models/nemotron-3-5-lightning) (a 30B-A3B MoE, 3.6B active) and [Gemma 4 26B A4B](/models/gemma-4-26b-a4b) (3.8B active) are exactly the kind of models a router hands the routine queries to, and [Llama 3.3 70B](/models/llama-3-3-70b-instruct) covers the mid-tier. AT&T’s 2% quality drop is the enterprise-shaped version of our own findings that mid-tier open models trade a few points for an order of magnitude in cost. ## The part AT&T is not saying out loud Every one of those models runs on a desk. Nemotron 3.5 Lightning’s NVFP4 build fits well under 32GB; Gemma 4 26B A4B runs on a 24GB GPU. An enterprise router paying per token for open-weight inference is paying a margin to someone else to host hardware you could own - and LiteLLM routes to a local Ollama or vLLM endpoint exactly the same way it routes to a cloud provider. If 40% of queries are routine enough to route away from the frontier, a chunk of them are cheap enough to run at home. Our [budget tool](/budget) does the own-versus-rent math for your volume, and the [rig finder](/find) matches a model list to the machine you already have. ## Context: the tokenmaxxing era is ending The backdrop to all of this: agent workloads consume far more compute than chatbots did, and the labs have shifted from flat subscriptions to token-based billing, so enterprise bills exploded. The June and July reporting PYMNTS cites describes companies building cost-management tooling and the end of “tokenmaxxing” - pushing employees onto the biggest models by default. AT&T is the flagship example of the counter-move: route aggressively, prefer open weights, hold frontier spend flat. **Honest framing.** These are AT&T’s internally measured numbers, reported secondhand through The Information - no benchmark suite, no per-task breakdown, and “quality” is AT&T’s own internal definition. The 56% ceiling applies to coding and select advanced tasks, not to the whole AI bill. But the direction, the mechanism, and the model families are all specific and consistent with independent evals, which is more than most enterprise AI savings claims survive.