The Model Routing Play
Corporate America is quietly ditching the idea that you need one single frontier model to run your company.
Data from Vercel shows that open-weight models jumped from 7% of platform tokens in December to 56% in August.
Even more telling: telecom giant AT&T now routes roughly 40% of its internal AI requests to open models like Llama, Gemma, and Nemotron. They only save premium frontier models for high-stakes work where errors are expensive.
On certain coding workloads, AT&T reported internal cost reductions of up to 56% while seeing only a tiny 2% drop in quality.
Why it matters
AI model performance is leveling out fast. A mid-tier or open model can now summarize customer calls, categorize inbox leads, or draft boilerplate code just as well as a flagship model.
Yet hundreds of businesses still pay premium pricing for routine, low-risk chores.
The competitive edge in AI has moved.
It is no longer about who has access to the newest foundation model. It is about building a smart traffic cop—an internal routing system that sends simple tasks to cheap models and complex tasks to premium models.
The opportunity
This creates two immediate ways to win:
For Business Owners (Direct Margin): If your company spends $10,000 a month on AI API calls, cutting that bill by 20% to 56% keeps $2,000 to $5,600 in your pocket every month. That is straight profit added back to your bottom line.
For Consultants and Agencies (New Revenue): You do not need to build foundation models. You can sell an AI Cost & Routing Audit to mid-market companies. You review their AI usage, set up task evaluation benchmarks, install routing rules, and take a monthly retainer or performance fee to manage their spend.
How you can use this today
Audit your current AI tasks: List every process using an AI model. Tag each task as low-risk (summaries, data extraction, tagging) or high-risk (legal checks, client-facing reasoning, complex code).
Create a test sample: Pull 25 real examples of your routine tasks (like recent support tickets or email inquiries).
Run a split test: Run those 25 tasks through a cheaper open-weight or mid-tier model and compare the output to your flagship model. Check if the cheaper output meets your business standard.
Install routing rules: Direct all basic tasks to the cheaper endpoint and reserve your premium API keys strictly for high-stakes tasks.