# AI Needs an Automatic Transmission

By Wes Pomeroy. Published June 29, 2026. https://wespomeroy.com/writing/ai-model-routing-to-cut-costs/

I upgraded my Anthropic plan today. That's not a brag. My car needed more gas, so I put more gas in it - the work I'm doing outgrew the plan I was on, and the upgrade was just the cost of keeping going.

At work it's the same story at a bigger scale. For a while the AI spend was small enough to wave through. It isn't small anymore. The line items add up, somebody now asks what we're getting for them, and "trust me, it's worth it" has stopped being an answer.

So I went looking at the part of this I'd been ignoring. And I found something that, a little embarrassingly, I should have set up before I ever upgraded anything.

## The Part I'd Been Skipping

Here's how most of us use these tools. You pick a model - usually a big, capable, expensive one - and you send everything to it. Easy questions, hard questions, "tidy up this sentence," "reason through this contract." It all goes to the same top-tier model, and you pay top-tier prices and top-tier wait times for every request, including the ones a much cheaper, faster model would have answered just as well.

There's a layer that sits between you and the model and fixes that. I've seen it called a few things - a gateway, a router, the efficiency layer - and the fuzzy naming is half the reason it stays invisible. What it does is easy to describe. It reads each request as it comes in, sizes up how hard the request really is, and sends it to the cheapest model that can handle it. When a request needs the heavy model, it goes there. When it doesn't, it goes to something smaller.

That one move saves money and time at once, for the same reason: the cheaper models are also the faster ones. Every request the layer sends down costs less and comes back quicker. On top of that, a good gateway trims the dead weight out of your prompts so you're not billed by the token for filler, and it reuses context it's already paid for instead of buying it again. Less money out, less time waiting, same answer.

## Why It Has to Be Automatic

Two reasons this doesn't happen on its own.

First, you're not going to do it by hand. Nobody stops before every question to decide whether this one deserves the expensive model or the cheap one. You'd spend more attention managing the tool than doing the work, and you'd guess wrong half the time anyway. The only version of this that survives a normal workday is the one you set up once and forget about.

Second, the model companies aren't going to do it for you. Their business is selling the expensive model. More usage on bigger models is the revenue. Asking them to route you down to something cheaper is asking them to make less money on purpose, quarter after quarter, in front of a board. No company built on usage is going to talk you out of usage. That's not a knock on them - it's just what they are. The push to save you money has to come from somewhere that doesn't bill you by the token.

## Getting It for Free

A lot of this is free, and the free version is good enough that it's where you should start.

The reference tool is [LiteLLM](https://www.litellm.ai/). It's open-source, it runs on a cheap server you already know how to rent, and it puts a single endpoint in front of more than a hundred models. You point your app at it instead of at OpenAI or Anthropic, set your own budget caps, and write your own rules for which requests go to which model. Nothing leaves your setup except the call to the model itself. And if you're already on Cloudflare, their [AI Gateway](https://developers.cloudflare.com/ai-gateway/) is sitting right there in your account - caching and rate-limiting included, at no extra charge.

If you're one person who just wants the benefit, you can build a stripped-down version in a weekend. A small, cheap model reads each request and guesses how hard it is. Easy ones get forwarded to a cheap model, hard ones to the expensive one. You strip the obvious junk out of long prompts. You switch on the prompt caching the model companies already offer and most people never turn on. None of these moves is hard on its own. Stacked together, they're most of what the paid products do.

## When You'll Want to Buy It

At some point the do-it-yourself version stops being worth your time. You've got a lot of people, you need audit trails, you want someone to call when it breaks. That's when you look at the managed gateways, and there are plenty. [OpenRouter](https://openrouter.ai/) gives you hundreds of models through one account with pricing you can read right off the page. [Portkey](https://portkey.ai/) open-sourced its whole gateway this spring and then got bought by [Palo Alto Networks](https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-to-acquire-portkey-to-secure-the-rise-of-ai-agents), who are folding it into their security platform - a decent sign of where this is going, since the layer is becoming something big vendors want built in rather than bolted on. [Helicone](https://www.helicone.ai/) leans on watching what your AI is doing; [Martian](https://withmartian.com/) focuses on the model-picking itself.

The one that got me looking in the first place was [Phantm](https://www.getphantm.ai/). It's a drop-in proxy that rewrites and routes your requests on the fly and claims to cut spend by close to half, with a running log of every change it makes so you can see exactly what it did. I'm not vouching for the number - that's the kind of claim you check yourself, which I'll get to. But it's the one that made the lightbulb go on for me, and it's a clean version of the pitch: same inputs, same outputs, smaller bill.

Across all of them, routing is becoming a standard feature, not a magic trick. Soon it'll either come built into whatever AI platform you already pay for, or you'll add a standalone layer for it. Either way, this stops being optional.

## You Can Test This, and You Should

What I like about this layer is that you don't have to take anyone's word for it.

Run a week of your actual work through one of these gateways, and run the same work the way you do it now. Two of the three things you care about you can measure flat out: cost shows up on your bill, and speed shows up on a stopwatch. No trust required, no vendor dashboard to believe.

The third thing - quality - you check the old-fashioned way. You put a person on it. You look at the outputs from the cheaper routing next to the outputs from your usual setup, on your own tasks, and you ask whether anyone can tell a difference that matters to you.

And that's where it goes past a simple cost cut. Once you can route by difficulty, quality stops being something you pick once and live with. It becomes a dial. You decide, job by job, how much horsepower the work is worth - spend big where the answer has to be excellent, spend little where "fine" is genuinely fine - and you run cost and quality as two knobs instead of one locked setting. For a small team that's a nice saving. For a company running thousands of these a day, it's a different way to run the operation.

## The Receipt in My Inbox

This is on my home list now, and it's going to stay in front of me - I'll look at it every time another receipt from Anthropic lands in my inbox. I'm raising it at work on Monday too.

The picture I keep is an automatic transmission. Nobody manually downshifts and upshifts their car anymore - and yet with AI, most of us are secretly redlining a lot of the time, running the engine flat out for a trip to the corner store. An automatic feels how hard the car is working and picks the right gear without you thinking about it. Picking the right model for the job should work the same way - automatic, invisible, quietly saving you money at the pump every mile.