Plan at the Frontier, Execute in the Trenches

The winning pattern isn’t “cheapest model everywhere” or “best model everywhere.” It’s frontier models for planning, and cheaper open-weight models for execution.

“I always start all of my work starting with the frontier… then I want to add sub-agents to them — super agents at certain skills.” Jensen Huang, NVIDIA

There’s a false choice at the center of most AI cost conversations: use the expensive frontier model everywhere and watch the bill explode, or use the cheap model everywhere and watch quality collapse. The teams that get this right refuse the choice. They use the right model for the right job — and the split is more intuitive than it sounds.

Start at the frontier for speed-to-clarity

Begin every hard problem with the smartest model you can get — the frontier model. Not because it’s cheap, but because it gets you to a correct plan fast. Frontier models are your senior partner: expensive per hour, but they scope the work, catch the traps, and hand you a clear plan of attack in minutes instead of days. Speed-to-clarity is worth the premium.

• Frontier model — the most capable models available at any moment (the senior partner).
• Open-weight model — a model whose parameters you can download, run, and customize on your own hardware.
• Sub-agent — a specialized worker an orchestrator delegates a narrow task to.
• Orchestration — the frontier model planning the work and routing pieces to the right sub-agents.

Hand off to the trenches

Once the plan is clear, the work itself — the repetitive execution — goes to cheaper, specialized models. And here’s the counter-intuitive part: cheaper execution can produce better answers, not just cheaper ones.

“When you have cost-effective intelligence, people just use more of it… you can iterate across a larger search space, and as a result the answer could actually be better.”

When each attempt is 10× cheaper, you can afford ten attempts where you used to afford one. You explore a bigger search space, compare more options, and keep the best. Cost-effective intelligence doesn’t just save money — it buys you more shots on goal. Open-weight models have gotten good enough to take those shots: on public benchmarks, NVIDIA’s open Nemotron 3 Ultra scores about 0.86 against a top closed model’s 0.87 — while running roughly 10× cheaper.

Build your routing table

The deliverable here is concrete: one row per class of task, mapping it to a model tier. This is the single artifact that turns “use the right model” from a slogan into an operating policy.

The pattern in one sentence

Plan with your best mind; execute with your cheapest reliable one; specialize where it pays. That’s the routing discipline — and it maps cleanly onto how you’d staff a team. You don’t put your principal on data entry, and you don’t put an intern on strategy. Onboarding your agents means giving each one the job it’s actually right for.