The One-Person Router: Stop Picking Your AI Model by Hand
The Departure
I caught myself doing it again last week, somewhere around the third gate change in Denver. A voice memo to clean up. An expense note to format. A proposal to tighten before a morning meeting. I fired all three at the most expensive model I had open, because it was the one already sitting there and I was too busy moving to think about it.
That habit used to be harmless. This week it got expensive to keep. In one eight-day stretch, the price of frontier-level AI dropped in three different places at once, and none of it helps you if you're still choosing your model by hand at a gate. The fix isn't another comparison chart. It's a rule that picks for you. Ramp built that rule for its 70,000 business customers over three years. You can build the one-person version in ten minutes.
The Co-Pilot
Tool: A ChatGPT Project called Model Router. (Same rule saved as Claude Project instructions, a Gemini Gem, or a Grok Automation works just as well.)
The Use Case: It decides which model handles a task the second you hand it over, so you stop defaulting every job to the priciest thing you have open.
The Build: In ChatGPT, open Projects, create one called Model Router, and paste this into its instructions:
"You are my model router. Below are the AI tasks I run most weeks, sorted by stakes.
Low-stakes (routine, internal, easy to check): [transcribing voice memos, formatting expense notes, summarizing an email thread, drafting an internal agenda].
High-stakes (client-facing, judgment calls, hard to walk back): [proposals, pricing replies, follow-ups on a live deal, anything a customer or exec reads].
When I hand you a task, match it to one bucket first. If it's low-stakes, run it on my cheapest capable model and keep it short. If it's high-stakes, say so and run it on my frontier model with full reasoning. Put one line at the top telling me the bucket and the tier you used, so I can override before you go further. If a task is genuinely unclear, ask me which bucket once, then remember my answer."
Now wire it to real tiers. Your cheap tier is for routine, checkable work: OpenAI's current published pricing now shows Luna at $0.20/$1.20 per million tokens, and Gemini Flash sits in the same neighborhood. Your frontier tier applies to anything a client reads: Opus 5 at $5/$25 or Sol at $5/$30. Anthropic's own numbers put Opus 5 at about 99.5% of Fable 5's capability on their coding benchmark at half the cost per task, which makes it a real frontier pick, not a compromise. Sonnet 5 sits in the middle at $2/$10, a strong default until that intro rate ends.
Want an open-weight option in the rotation? The news this week wasn't the 2.8-trillion-parameter download. You can't practically run that on your own hardware. It's that the same model is hosted and ready at $2.90/$14 per million, an API key, not a server in your closet.
Here's why this earns its keep on the road. Once the rule lives in the Project, every task you fire from a gate gets sorted without you thinking about it, and any scheduled Codex task that runs while you're in the air inherits the same routing. The decision gets made once, at your desk, and travels with you.
The Upgrade
Topic: Build the one-person router
List your week, then tag it. Write down the 5 to 8 AI tasks you actually run most weeks. Mark each one low-stakes (routine, internal, easy to check) or high-stakes (client-facing, hard to walk back). The tags are the whole system.
Paste the rule where you already pay. Drop the router instruction into a ChatGPT Project, a Claude Project, a Gemini Gem, or a Grok Automation. Pick one. You don't need a new subscription. You need the rule saved somewhere it runs every time.
Put August 31 on your calendar. Sonnet 5's intro rate ($2/$10) ends that day. Set a reminder to re-check your cheap tier that week so a price change doesn't quietly move your default back up.
The Landing
Ten minutes today, wherever you're sitting. Open the assistant you already pay for, list your recurring tasks, tag them as low- or high-stakes, and paste the router rule. Then fire your next routine job at the cheap tier on purpose, rather than out of habit. That's the whole shift: you stop deciding which model to use, and start letting the rule decide while you keep moving.
Safe Travels,
Marcellus