Blog

The moat moved upward. Here's the layer it moved to.

When an open model ties the frontier at a third of the price, the advantage stops being which model you own and becomes the system that decides how, when, and where to use each one, with the evidence to prove it. That system has a shape, and most teams are describing it without building it.

July 20, 2026 · 5 min read

An open-source model just scored within a few points of the frontier on the public intelligence index, topped the coding benchmarks outright, and bills roughly thirty cents on the dollar. The predictable reaction is a spreadsheet debate: is the last handful of intelligence points worth paying four times as much? That's the wrong question, and the more interesting answer is already circulating in the replies to every post about it.

One operator put it precisely: the moat moves upward, from owning the smartest tool to designing the organization that knows how, when, and where to use each one. Route work by complexity, risk, cost, and required verification. Use the inexpensive model where it suffices. Escalate the hard reasoning when the stakes justify it. Keep evidence and review gates consistent across both. That is not a hot take. That is a system specification, and a lot of smart people are writing it down without noticing they've described a product that has to exist.

The specification, and why it's harder than it sounds

Say it plainly and it sounds obvious: cheap model for the easy 90%, frontier model for the hard 10%, and something decides which is which. The difficulty is entirely in that last clause. Deciding per request, automatically, without a human triaging every prompt, is the part that turns a good idea into an afternoon script that quietly caps out. Matching keywords to a model tops out around a third accurate, because "fix this" can be a typo or a concurrency bug and the words don't tell you which. The decision has to estimate difficulty, not scan for vocabulary.

And the escalation half has a trap the cost conversation skips: how do you know the cheap model's answer actually held up? "We downrouted and saved money" is a claim. If the cheap answer was quietly worse, the savings are a quality regression wearing a discount. A system that routes by cost without measuring whether the cheap answer held is optimizing the number it can see and ignoring the one that matters.

The half everyone leaves out: evidence

Here's where the LinkedIn version of this thesis stops and the regulated-industry version keeps going. "Keep evidence and review gates consistent across both models" is easy to type and hard to build. When a request could go to a cheap open model from one jurisdiction or a frontier model from another, the questions multiply: which model handled this, and why? Was the prompt screened before it left? Can you prove the record wasn't edited after the fact? Which of these models is even allowed for this workload? Owning the routing decision is worthless to a bank if you can't produce the evidence of what the routing decided and did.

This is also where the cheap-model story collides with a real constraint people are naming out loud: several of the strongest low-cost models are built in high-risk jurisdictions. The answer for a regulated firm isn't to ban them, it's to know which model ran, govern whether it's permitted, and record it. Cheap and governable are the same decision, made once, or they don't happen together at all.

So that's the layer, and it's what RemKey is

The system the market is describing is the AI layer you can prove to a regulator, and it's one base-URL swap. A classifier estimates each request's difficulty and routes it to the cheapest capable model (the accuracy, the sample sizes, and every miss are published on the benchmarks page, because a routing claim you can't inspect is just a number). A sampled check runs the frontier model behind the scenes and scores whether the cheap answer held, so the savings come as a verified quality-hold number instead of a hope. Every call is screened for PII and injection before it leaves, fail-closed. Every model carries its origin and jurisdiction in the catalog so your policy can allow or block it on your terms. And all of it lands on a hash-chained, Ed25519-signed ledger your auditor verifies offline, without trusting us. Routing, escalation, and evidence, consistent across every model, which is exactly the specification, built.

The frontier will keep getting cheaper, and every drop makes this truer, not less: when the models converge, the entire advantage is the system that decides among them and can prove what it did. Owning the smartest model was never going to be the moat. Owning the governed decision is. Swap the base URL and the decision is governed from the first request.

See how the routing decision is measured What this looks like under a compliance review