DevRev's new autonomous support stack on GPT-6.1 Sol shows the layering any small business should copy: a cheap fast model handles routing and simple tickets, a reasoning model only kicks in on hard cases. The lesson for a 50-ticket-a-day shop is to split the pipeline by difficulty before buying any premium model tokens.
Per OpenAI, DevRev powers its Computer and Operator agents with GPT-6.1 Sol, balancing reasoning, speed and cost inside one production service stack. Simple refund and order-status queries stay on fast cheap models; multi-step diagnostics and account actions escalate to Sol. The result is self-service that doesn't burn a reasoning token on every 'where is my order'.
What it means for a small business
You don't need enterprise scale to use this pattern. In a Claude or GPT support flow, a lightweight classifier (Haiku, 4o-mini, or a cached intent prompt) tags intent and risk, and only the hard tickets hit a reasoning model. For a shop on WhatsApp or email, that typically cuts model spend 5-10x while keeping quality where it counts - the same shape I use in client service automation for a fitness network, where booking confirmations and genuine complaints travel through different lanes.
Why it matters: Autonomous support agents make economic sense only when you split cheap and expensive thinking - route first, reason only when the ticket earns it.
Source: OpenAI