Three decision models route requests for about a hundredth of a cent

TLDR

Since September 15, three decision models shipped: Jev (TypeSafe), OpenJev (TheoLeeCJ), and Laya (ConvAI Innovations). Each one answers a question with a fixed set of possible answers and a confidence score for roughly a hundredth of a cent. Use one when you need to pick a model per request and do not want a hosted router. Five companies raised venture money selling that same routing judgment as a product. The routing function is now a small model.

Hosted routers sold one decision: which model handles a request

Between 2023 and 2026, the model-routing layer raised real money on one promise: send each request to the right model, cheaply.

LiteLLM: $1.6M seed for the open-source gateway
Martian: $9M for a router that claimed to beat GPT-4 at picking models, with Accenture investing
NotDiamond: $2.3M for model routing
Unify: $8M for LLM deployment optimization
OpenRouter: $40M raised at a $5 billion valuation, the routing layer as marketplace

The total was about sixty million dollars against one job: decide which model handles a request. If a vendor charges per request for that decision, compare it with Jev, OpenJev, or Laya before renewing.

How to route a request in one pass

A decision model does that job in one pass. The request and the fleet description go in: this model is fast and cheap, that one is good at code, this one handles vision. Back comes a model name and a confidence score. No essay. No parsing step. No retraining when the fleet changes.

The cost of one routing decision is a few hundred tokens. Four hundredths of a cent on Jev. Nothing on Laya, self-hosted. Twenty-five thousand routing decisions cost a penny - the same judgment the venture-funded routers bill per request or per seat.

What you still maintain when you run a decision model

A decision model does not know a model’s capabilities unless you tell it. The fleet description is yours to maintain: which models exist, what each is good at, what each costs. It does not benchmark arrivals, track live pricing, or learn from a wrong pick. New frontier model lands overnight? You add it to the list yourself.

It also cannot do what the gateway products layer on top: billing across sixty providers, key management, SLAs, audits, dashboards. OpenRouter is worth five billion dollars because of those gateway layers, not because of model selection. Jev is not its replacement. Jev replaces the part that called itself intelligent routing.

Why you no longer train a classifier to pick models

The routing companies existed because the alternative was worse: either pay frontier-model prices for every request, or build and maintain a trained classifier per task. That meant labeled data, ML engineers, GPUs, and a retraining cycle every time the categories moved.

A decision model removes the training cycle. The answer set is defined at the moment you ask. Laya’s author published the underlying RL research in March 2025; Jev productized it with the right interface at the right pricing moment; OpenJev and Laya proved the idea runs in a browser tab and in a hundred languages. Three teams independently shipped the same capability in one week. The products that charged for that capability lost their pricing power.

Which vendors lose if you run routing yourself

Pure routing plays (the router was the whole product): fully exposed. Their moat was access to model metadata and a scoring function; both are now commodity.
Gateways and marketplaces (OpenRouter, LiteLLM): the routing feature is eaten, the business is not. They still normalize APIs, bill, and broker. Expect them to ship decision-model routing themselves, and OpenRouter is already the natural place to run Jev for exactly this.
Labs’ built-in auto modes: ChatGPT and Claude route internally. A closed model that also sells you the classifier for your own fleet is an interesting position to watch.

Routing vendors once charged for model selection because it required a trained classifier. That classifier now runs on a T4 or in a browser tab. Model routing is now a primitive you can run yourself.

How to route local open models for free

Laya runs on a T4. OpenJev runs in a browser tab. The router that venture capital funded as a hosted product now runs on a laptop or a phone, in a hundred languages, for free. If you run open models behind your product, put Laya or OpenJev on the same machine and pay nothing per routing decision.

Sources: Adam Azzam on Jev as a model router (Modal), Laya model card, SalesRLAgent, March 2025, funding pages for LiteLLM, Martian, NotDiamond, Unify, OpenRouter, Jev guide on Tokenstead.

Sources verified September 19 to 20, 2026. Funding figures are as reported by the linked outlets.