Analysis
AT&T is expanding its use of open-source AI models specifically to reduce what it pays Anthropic and other proprietary-model providers, The Information reported. AT&T processes roughly 45 billion tokens a day across its AI operations, a volume large enough that routing every query through frontier, generalized models becomes, in the company's own framing, financially unsustainable -- prompting AT&T to build a customizable, open-source-based routing system instead.
The company built an internal AI gateway that uses cache-aware routing to match each incoming task to the most cost-effective available model without sacrificing output quality. The system can switch models partway through a multi-turn conversation, weighing speed, cost and expected quality at each stage before routing a given prompt to whichever model -- open or proprietary -- makes the most sense for that specific step. AT&T says it's expanding open-weight model usage from about 25% of its AI operations today to as much as 80%, and that the shift has already produced savings of 80% to 90% versus proprietary models in certain applications.
Part of a broader efficiency shift
AT&T's move fits a pattern CNBC identified earlier this summer, describing how OpenAI and Anthropic face a new reality as enterprise users shift from "tokenmaxxing" to efficiency -- the early-2026 pattern of routing every task through the most capable, most expensive frontier model has given way to more disciplined, cost-aware routing as enterprise AI bills have scaled with usage. AT&T's case is a particularly aggressive version of that shift given the sheer token volume involved: at 45 billion tokens a day, even small per-token cost differences compound into meaningful budget line items, which is exactly the kind of scale where building custom routing infrastructure pays for itself relatively quickly.
AT&T's framing -- that open models produce "more meaningful, telco-specific outcomes" -- deserves some scrutiny alongside the cost story: open-weight models generally require more in-house fine-tuning and infrastructure investment to match proprietary-model quality on specialized tasks, meaning some of the reported savings may be shifting cost from AI vendor bills to internal engineering headcount rather than eliminating it outright. AT&T has not disclosed the internal engineering cost of building and maintaining its routing gateway, which is the necessary other half of a genuine cost comparison against simply paying Anthropic's or OpenAI's list prices.
