Everyone's watching the AI price war for the wrong reason. The headlines are about how cheap tokens got. The actual story is what cheap tokens do to how you should be architecting software right now.
Here's what changed, and why it matters more than the price cut itself.
The major labs have quietly split their lineups into tiers. Cheap, fast models for routine work. Expensive, deep-reasoning models for the hard stuff. This isn't a pricing gimmick, it's an architecture signal. If your app sends every request to the same model regardless of difficulty, you're either overpaying for simple tasks or underpowering the hard ones.
The pattern worth adopting: route by task, not by app.
function routeModel(task) {
if (task.type === 'classification' || task.type === 'extraction') {
return 'cheap-tier-model';
}
if (task.type === 'reasoning' || task.type === 'multi-step-planning') {
return 'frontier-tier-model';
}
return 'mid-tier-model';
}
Simple idea. Almost nobody's actual codebase does this yet. Most apps still hardcode one model for everything, which made sense a year ago when the price gap between tiers was small. It isn't small anymore.
Multiple frontier models now ship with million-token context windows. A year ago, retrieval augmented generation existed mostly because you had no other choice, you couldn't fit enough context in the window, so you chunked, embedded, and retrieved. That constraint is disappearing fast for a lot of use cases.
This doesn't mean RAG is dead. It means the decision of when you actually need it just got a lot more deliberate. If your dataset fits in context, a vector database might now be solving a problem you don't have anymore.
While the pricing race gets all the attention, the EU AI Act's high-risk provisions became enforceable this month. Transparency rules now require chatbots to identify themselves as AI, and synthetic media needs to carry labels. If you're shipping anything AI-facing into the EU market, this isn't optional anymore, and I'd bet most side projects and even a few production apps aren't compliant yet.
A year ago, the interesting AI engineering question was "which model." Now it's "which model, for which task, at which cost, under which rules." That's a genuinely different architecture problem than the one most tutorials are still teaching.
Is anyone here actually doing tiered model routing in production, or is everyone still hardcoding one model per app? Curious how far ahead or behind the rest of us actually are on this.