Riften routes each AI request a company makes to the lowest-cost model that can handle it, evaluates the results against the outcomes the company actually cares about, and helps that company build models it owns.

That is the product. The argument of this piece is about what it sits on top of: the valuable thing is not the model, it is the moment a person decides where to type.

What they built

Riften is an AI gateway. You connect the model providers you already pay for, and it sends each request to the cheapest model actually capable of handling it, so you stop paying frontier prices for work a smaller model does perfectly well. It checks the work too, scoring results against outcome-based evaluations rather than guesswork. And where a company chooses to enable content capture, Riften can use that company's own traffic, judged by those evaluations, to help it train models the company owns.

Carl Li He Stauffer, founder and chief executive, was a forward deployed engineer at Palantir and a nuclear researcher at Jefferson Lab, studied computer science and applied mathematics at Johns Hopkins, and was an a16z Scout. The rest of the team comes from Stanford, Berkeley and Carnegie Mellon, and from Amazon Search, Palo Alto Networks, UCSF Health and NASA. The site is riften.ai.

Standing between a user and a model is a crowded place. OpenRouter made routing across providers ordinary. Portkey, LiteLLM and Helicone sell the gateway to engineering teams. Perplexity and Poe went after the consumer typing box directly. Further out, Experiential Labs is building the layer that decides what kind of intelligence a request deserves, and Turnstone is assembling the context an agent needs before it answers. Every one of them is reaching for the same thing from a different direction.

The point: this is a fight for the prompt

Put it simply. There is going to be a long run of startups whose real business is fighting for the human's prompt.

Riften has found a good reason to hand yours over: it looks at each request and routes it to the cheapest model that can do the job. That is smart, and it is honest about who it is for. A company with a large monthly bill cares enormously. A person typing three questions an evening does not care at all.

So the reason is not the opportunity. The prompt is the opportunity, and a reason is just how you get it. Any startup that finds a persuasive one, for individuals or for companies or for agents, is in this business.

It is worth being precise about why a prompt is worth this much. It is not only that the person came to you rather than someone else. It is that the request itself says what people actually want done, at what price, and what a good answer looked like. Where a customer enables content capture, Riften can turn exactly that, scored by outcome-based evaluations, into a model the customer owns, which is the clearest possible demonstration that the flow is the asset and the routing is the excuse.

Why the model companies end up buying

This section is editorial speculation on our part. It is not something Riften has said, and not a description of its plans.

Anyone holding a meaningful flow of prompts is a natural target for the companies with the most money, and those happen to be the model companies, who are in an expensive fight for users right now.

There is an obvious objection, and it deserves answering rather than skipping. Why would a lab buy you instead of simply crushing you? It can cut prices, or ship its own router in a fortnight.

Because what you hold is not the software, it is the habit. A lab can copy a routing engine easily and cannot copy, to take a purely hypothetical example, ten thousand companies already pointing their traffic at you, have signed with you, and are not thinking about it any more. Distribution is the slow thing. Buying it is faster than winning it, and in a land grab, speed is what the money is for.

That is also why the feature set matters less here than it usually does. It is the audition, not the asset.

The reasons never run out

This is opinion rather than anything the company has said, and it is the part worth keeping.

The best thing about this category is that it renews itself. Every month and every year, somebody finds a new reason why you should type into their box instead of the one you used last year.

Cheaper is this year's reason. Private is next year's, for everyone who cannot send work to an American lab. Then it is the one that remembers you, the one that never loses your files, the one cleared for medical use, the one that works offline, the one your regulator has already approved, the one with taste, the one your whole team can see, the one that answers in your own language properly, the one that is simply faster.

Each of those is a company. None of them needs the others to fail. And unlike most categories, being late is survivable here, because the reason that wins in 2029 has probably not been thought of yet.

These opportunities are never-ending, and that is unusual enough to be worth saying out loud.

Why it is bigger than it sounds

The short description is a cost-cutting gateway. What it is competing for is much larger than the saving.

  • The prompt is the prize, not the routing. The reason you give somebody is how you get the flow, and it can be almost anything as long as it persuades.
  • Holding the flow could make a router worth buying. This is our speculation: not for the software, which is copyable, but for the habit, which is not, and the labs are in a hurry.
  • New reasons arrive every year. Cheaper, private, remembers you, approved, offline, faster. Each one is a fresh company and being late costs less than it usually does.

What to watch

The measure is how much traffic runs through it, not how much money it saves. Savings are the pitch and volume is the position.

The second is how many customers enable content capture and adopt models of their own trained on it. If they do, Riften stops being a router and becomes a supplier, which is a considerably better place to stand.

The third is what happens at the first big price cut from a frontier lab. A company whose only reason to exist is cost has its reason set by somebody else, and this is the moment that shows whether the habit is real.