AT&T cut the cost of AI used for coding tasks among its employees by 56% — not by switching vendors, but by automatically routing most routine queries to cheaper open-weight models and reserving Claude and GPT for what actually needs frontier-level capability. The measured quality drop was just 2%.
What changed
According to The Information, PYMNTS, and CryptoBriefing, AT&T implemented model routing via LiteLLM to direct routine coding queries — the bulk of daily volume — to cheaper open-weight models, reserving Anthropic's and OpenAI's frontier models for tasks that need more capability. The setup already runs at production scale: roughly 150,000 employees, 45 billion tokens processed daily, with about 40% of usage currently on open models and an expectation of reaching 60-70%. Between February and July, the company also tested its own telecom-tuned models, which delivered up to 90% savings in inference cost for industry-specific tasks. AT&T's stated goal is to keep spending with Anthropic and OpenAI flat, not growing, as routing to open models increases.
Why it matters
AT&T's case is the practical answer to exactly the problem we've covered here over the past few weeks: DeepSeek's repricing, Gemini's scheduled discount expiration, the financial contrast between OpenAI and Anthropic. If any AI vendor's price can change without notice — because its business model hasn't closed the loop yet — the structural response isn't betting everything on whichever vendor is cheapest today, it's building a routing layer that decides, task by task, which model meets the job at the lowest acceptable cost. The 2% quality drop shows that, for most routine tasks, the more expensive model is capability the company was paying for without needing it.
The impact for Brazil
For Brazilian companies already running AI at scale and feeling the weight of per-token cost — especially after the repricing episodes we've covered here recently —, AT&T's model is replicable without needing telecom-carrier scale: the logic of separating routine tasks (which tolerate a cheaper model) from critical ones (which need frontier capability), and automating that decision through routing, works at any volume. It also reduces dependency on a single vendor — the same concentration risk we discussed covering SpaceX's acquisition of Cursor and the geopolitical dispute between Pax Silica and WAICO.
Entercast's take
This is a direct counterpoint to NVIDIA's NOOA, which we covered here earlier this month: efficiency gains can come from agent architecture, or they can come from an intelligent routing layer across models — they don't have to come from the most advanced model available. Before signing another expensive contract with a single AI vendor, it's worth measuring how much of the company's current usage is, in practice, routine work that a cheaper model would handle at an acceptable quality cost. The answer may be worth more than any promotional discount.