Business

AT&T Cut Coding Costs 56% by Switching to Open Models.

A tweet spread saying AT&T routes 40% of employee AI usage to open models and cut coding costs 56%. The person who made the remark was the VP of Data Science, not the "AI chief" the tweet named.

· Huuush

AT&T Cut Coding Costs 56% by Switching to Open Models

Key takeaways

  1. 1AT&T said it routes 40% of employee AI usage to open models and plans to raise that to 60-70%, with coding costs down 56% and quality down only 2% (cross-checked against AT&T's official blog and PYMNTS).
  2. 2The viral tweet attributed the remark to the "AT&T AI chief," but the screenshot attached to it quotes Mark Austin, VP of Data Science. AT&T's official Chief Data and AI Officer is a different person (Andy Markus).
  3. 3AT&T still uses Anthropic and OpenAI frontier models for demanding tasks like code generation and moves only relatively lighter tasks to open models like Nvidia Nemotron, Meta Llama, and Google Gemma.
Screenshot of an article quoting AT&T on open models, with the phrases 'open source' and 'just as good or better' highlighted

A tweet posted by X account @Hesamation in the early hours of August 21 (KST) spread, drawing more than 170,000 views. It said, "AT&T is running 40% of employee AI usage on open models and plans to raise it to 60-70%" and "coding costs down 56%, quality drop only 2%, 45 billion tokens processed a day."

45 billion tokens a day: AT&T's open-model bet

These figures aren't just one tweet's claim. PYMNTS and AT&T's official blog cross-confirm the 56% coding cost reduction, the 2% quality drop, and the goal of raising open-model share from 40% to 60-70%. The 45-billion-tokens-a-day throughput is also confirmed by a TelecomLead report. The original source is the Applied AI newsletter from the paid outlet The Information, titled "AT&T is Using Open Source Models to Curb Anthropic Bills."

MetricFigure
Share routed to open models40% → target 60-70%
Change in coding costs-56%
Quality drop-2%
Daily token throughputAbout 45 billion tokens

The "AI chief" remark was actually someone else

The viral tweet attributed the statement that "open models are just as good or better for many tasks" to the "AT&T AI chief." But opening the screenshot actually attached to that tweet, the speaker is Mark Austin (VP, Data Science). AT&T's official Chief Data and AI Officer is a different person, Andy Markus. The words of two people with different titles got lumped together into one tweet line.

Code generation is still for frontier models

AT&T hasn't switched entirely to open models. It's a two-track strategy: it still uses frontier models from Anthropic and OpenAI for complex tasks such as code generation, and moves only relatively lighter tasks such as code summarization to open models like Nvidia Nemotron, Meta Llama, and Google Gemma. It also uses OTel 2.0, its own fine-tuned version of Google Gemma 3 31B-IT, and the DeepSeek and Moonshot families are only being evaluated and haven't been adopted yet. For companies designing their AI cost structure, this approach of splitting by task difficulty, rather than "all open models" or "all frontier models," is worth referencing.

Frequently asked questions

How much did AT&T save with open models?
Coding costs fell 56% and quality dropped only 2%, according to the reporting (cross-checked against AT&T's official blog and PYMNTS).
Who said that the "AI chief says open models are better"?
The tweet that first spread attributed it to the "AT&T AI chief," but according to the screenshot attached to the tweet, the actual speaker is Mark Austin, VP of Data Science. AT&T's official Chief Data and AI Officer is a different person, Andy Markus.
Which open models does AT&T use?
It uses Nvidia Nemotron, Meta Llama, Google Gemma, and its own fine-tuned OTel 2.0. DeepSeek and Moonshot are being evaluated and haven't been adopted yet.
Does it stop using Anthropic and OpenAI models entirely?
No. It still uses frontier models for complex tasks like code generation and moves only relatively lighter tasks such as summarization to open models.

Related articles