AT&T Cut AI Coding Costs 56% With Model Routers — Without Killing Quality

AT&T is processing 45 billion AI tokens per day — up from roughly 8 billion a year ago — and its overall AI spend is relatively flat. That is not a procurement miracle; it is a routing architecture. On August 20, 2026, The Information reported that VP Mark Austin, who oversees employee-facing AI at AT&T, cut coding and advanced-task costs by up to 56% using LiteLLM-based model routers, with only about 2% quality degradation.

For engineering leaders watching Databricks' four cost levers or debating whether to give every developer unlimited Claude Code access, AT&T's numbers are the enterprise case study that turns "model routing" from a blog buzzword into a budget line item.

TL;DR: what people are asking

The tokenomics strategy in three moves

1. Cache-aware routing (the 56% lever)

AT&T's gateway evaluates each request on:

• Task complexity — can a cheaper model handle this?
• Model pricing across providers
• Cache state — cached tokens cost ~one-tenth of uncached

The counterintuitive insight Austin shared with Fierce Network: on long agentic coding sessions, staying on a premium model can be cheaper because the entire conversation context is already cached. Switching models mid-session throws away that cache.

This is the same long-running-task cost dynamic Claude Code auto mode touches — but AT&T operationalized it at 45B tokens/day.

2. Open-model substitution (the 80–90% lever)

AT&T audits deployed closed-source models older than a year and asks: can an open model handle this with a tweaked prompt? Austin estimated ~25% of total token volume fits that profile — network management, classification, simpler generation.

The company released OTel 2.0, an open-source telco-specific model, and trains additional open models for accuracy on network operations tasks.

3. Flat spend on rising volume

Token volume grew 5.6× (8B → 45B/day) while spend stayed flat. That is the outcome routing plus substitution is designed to produce — and it is what optimising generative AI costs looks like when it stops being a slide deck and starts being production infrastructure.

What you can copy without AT&T's scale

How this fits the broader enterprise playbook

AT&T's story validates the same architecture Databricks documented a week earlier with input from Stripe, Coinbase, Uber, and Ramp:

• Efficiency frontier (best price for a quality bar) matters more than intelligence frontier (peak capability) for everyday coding.
• Hard token cutoffs are a last resort — they punish your most productive developers.
• Context window pricing and cache hit rates are first-class routing inputs, not afterthoughts.

Related on explainx.ai

• Databricks: managing AI coding costs at scale — four levers with Stripe/Coinbase input
• Optimising generative AI costs — explainx.ai's cost framework
• Context window pricing decoded — why cache hits change routing math
• Claude Code pricing guide — model tier costs
• Loop engineering with Claude Code — long-running task patterns
• Choosing open-weight vs closed models — when substitution works

Sources: The Information — AT&T AI cost routing (Aug 20, 2026) · Fierce Network — AT&T tokenomics · Telecom Ramblings — tokenomics equation

Figures attributed to Mark Austin and AT&T leadership are accurate as of August 20, 2026 reporting. Routing implementations vary by organization — validate savings with your own eval suite.