News analysis · Published
AT&T's Open-Model Shift: How Businesses Should Route AI Workloads
By the ELYMENT AI editorial team · Free to read
AT&T says open models now handle 40% of its AI use, up from 20% in May 2026, and could reach 60% in coming months. The lesson is not that every business should replace frontier models. It is that AI estates need an evidence-based routing layer: send each task to the least expensive approved model that still meets quality, latency, security and reliability requirements, then escalate only when the measured outcome justifies it.
What changed in AT&T's model mix
The New York Times reported on 6 September, via The Indian Express, that open models accounted for 40% of AT&T's AI use, compared with 20% in May. Chief Data and AI Officer Andy Markus said the share could rise to 60% in coming months and that the company was saving up to 80% on AI costs compared with earlier in 2026. Those are AT&T's reported results, not a universal saving rate.
AT&T's own 23 July technical account describes an average workload of 45 billion tokens a day. Its cache-aware AI Gateway weighs expected output quality, speed and cost at each turn, and AT&T says the approach has reduced costs by as much as 90% in some settings. The two percentages use different comparisons and should not be blended into one benchmark.
Open models are an option, not an outcome
An open-weight model can reduce licence or API charges, support deployment on controlled infrastructure and allow specialisation. It can also create new costs for serving, monitoring, security, upgrades and engineering. A zero-price download is not a zero-cost production system.
The useful unit of comparison is therefore an accepted business outcome: a correctly classified ticket, a reviewed code change or a resolved service request. Compare total cost, completion quality, latency and intervention rates for the same task. Frontier capability may remain justified for difficult reasoning, while a smaller open model can be the better operational choice for stable, narrow work.
Build a workload routing scorecard
Start with a portfolio view rather than a vendor preference. For each production workflow, record:
- Outcome quality: task-level acceptance criteria, failure severity and the evidence required before release.
- Economics: input, output and cached-token cost plus hosting, orchestration, review and support overhead.
- Operational fit: response time, context length, tool use, throughput, availability and recovery behaviour.
- Control boundary: data location, retention, model licence, supplier terms, logging and permitted downstream use.
- Escalation rule: the observable conditions that move a request to a more capable model or human reviewer.
Routing needs its own controls
A router can quietly change which supplier or model processes sensitive work. Treat its policy as production configuration: approve the model pool, pin versions, preserve the routing decision and cost record, and test fallbacks. Cache-aware decisions matter because switching models during a long conversation can discard discounted context and erase expected savings.
Fierce Network reported in August that AT&T scores routing decisions on thousands of benchmark tasks and reviews older use cases for migration. That is the transferable pattern. Re-evaluate mature workflows periodically, but do not lower the acceptance threshold merely to increase the open-model share.
What business leaders should do next
Rank current AI workflows by spend, volume and consequence. Select two high-volume, lower-risk tasks; benchmark an approved smaller or open model against the incumbent; include review and infrastructure costs; and define an automatic escalation path. Promote only when the alternative meets the same outcome threshold across representative production cases.
ELYMENT AI's guides to open-weight deployment, private AI responsibility and model-upgrade evaluation provide the surrounding procurement and testing questions. ELYMENT AI can help teams turn that evidence into a governed routing policy and a cost-per-accepted-outcome dashboard.
Sources
- AT&T: The Tokenomics Equation (23 July 2026) - Primary account of AT&T's cache-aware AI Gateway, 45-billion-token daily workload and provider-reported cost results.
- The New York Times via The Indian Express: Corporate America is getting hooked on open-source AI (6 September 2026) - Current independent reporting on AT&T's open-model share, stated near-term target and reported cost reduction.
- Fierce Network: Open models are driving AT&T's AI tokenomics strategy (18 August 2026) - Independent interview covering AT&T's use-case review, routing benchmarks, escalation and cache-aware cost decisions.
- The Wall Street Journal: Why AT&T is betting big on open-weight AI (12 August 2026) - Independent reporting on AT&T's earlier open-model share and its focus on token cost and proprietary-data control.
Continue learning
Related analysis
- RoboParty RP1: What Businesses Need From an Open Robotics Release
RoboParty's RP1 roadmap highlights a practical buying test: verify available components, reproducible builds, licence scope and maintenance before investing.
- BRICS AI Open-Source Plan: What Businesses Should Verify
China proposed a BRICS AI open-source community on 13 September. Learn the licence, provenance, security and portability checks businesses need.
Frequently asked questions
Does AT&T use only open AI models?
No. Reporting on 6 September 2026 said open models handled 40% of AT&T's AI use. Its strategy combines open and proprietary models and routes work according to expected quality, speed and cost.
Do open models automatically reduce AI costs?
No. Savings depend on the workload, hardware, serving efficiency, cache behaviour, engineering effort, review burden and whether the model meets the required outcome quality.
What should an AI model router record?
Record the task, selected model and version, routing reason, relevant data boundary, latency, token or infrastructure cost, output result, escalation and human intervention.