Technology & Machines 👩‍💻⚙️

Big AI companies are now in trouble These researches just exposed a massive leak in how we use AI It’s called “the frontier tax.” And it explains why your AI agent is quietly burning money on almost every step it takes. Right now, agents run for hours. Task lengths went from ~11 minutes in 2024 to likely 16+ hours in 2026. But there’s a massive problem. Most of those hundreds of steps don’t need the smartest model. Renaming a file and redesigning your architecture get billed at the exact same frontier price. What happens when an AI picks the right model for every single step? A new technical report from Rayline just gave us the answer. And it is wild. Their tiny 0.8B model, ARC, predicts which model will succeed at each step and what it will actually cost, then picks. Across six popular benchmarks, ARC hit 97.5% of Opus 5.5’s quality at up to 92.9% lower cost. On DeepSWE, it scored 68.7% vs Opus 5.5’s 69.9%... for 7% of the cost per task ($0.18 vs $2.61). It’s like paying a senior engineer’s rate for every typo fix. Every step gets its own decision: a small model for the easy stuff, a frontier model with high thinking for the hard stuff. And it stays put when switching would cost more. For years, we thought the way to get better AI results was to use the biggest model for everything. This report proves the real bottleneck is picking the right model. If we keep running every step on the most expensive model, we’ll keep paying Big Token for intelligence we don’t need. The ultimate irony? We thought the fix for expensive AI was a cheaper model. Instead, this research proves the fix is a tiny model whose only job is choosing the model.