Blog
Router, gateway, blended model: the terms, untangled
The category's vocabulary is a mess and vendors benefit from the blur. Here are the three layers defined precisely, and the one question that separates them: who picks the model for each request?
A gateway is the integration layer: one API surface for many providers. A router is a gateway plus selection: rules or logic decide which model serves each request. A blended model moves selection inside the request itself: several LLMs run on every request and one synthesized answer returns, billed at cost in Pareto's case. The separating question is who picks the model.
Gateway: the plumbing
Every product in this category is at least a gateway: your code speaks one API (in practice, the OpenAI-compatible surface) and the layer speaks to every provider. LiteLLM self-hosted, Portkey hosted, OpenRouter's catalog, all gateways first. The gateway solves integration and changes nothing about decisions: you name the model on every request.
Gateway shopping is therefore fee-and-operations shopping: 5.5% on credits, 5% markup, $49 flat, or free plus ops. We keep the full math in the fee comparison.
Router: plumbing plus a decision-maker
Routing adds selection logic on top: fallbacks when a provider fails, rules like 'short prompts go to the cheap model,' provider steering like Exacto. The decisions follow a policy someone wrote, and that someone is you: routing rules are a config file with your name on it, re-litigated at every price event and launch, three of which hit the flagship tier in the last nine months.
Routing is genuinely valuable; the fine print is only that the intelligence is static. The rules know what you knew when you wrote them.
Start from your constraint
Honest by design: two of the four answers are not us. Fees verified 2026-07-28; the deeper walkthrough of each trade is above.
Blended model: the decision moves inside
A blended model stops routing between models and starts operating as one: every request runs several LLMs, the results are synthesized, one answer returns. There is no model parameter, no routing table, and no homework when the market reprices. That is the category definition, and Pareto is ours: billed at cost, benchmarked against Opus 4.8 with published receipts, up to 3x slower on the hardest reasoning tasks and +0 on agentic work, disclosed rather than discovered.
The three layers are not rivals so much as answers to different questions: integration (gateway), policy (router), and delegation (blended model). Decide which question you are actually asking; the vocabulary sorts itself out after that.
Questions, answered
Is OpenRouter a router or a gateway?
Both, by these definitions: a gateway (one API, many providers) with routing features on top (fallbacks, provider steering like Exacto). Model choice remains yours, which is what separates it from a blended model.
Is a blended model just automatic routing?
No. Routing picks one model per request from rules. A blended model runs several LLMs on every request and synthesizes one answer; selection stops being an event and becomes part of inference itself.
Which layer should I buy?
Match the question: need integration only, buy the cheapest gateway your ops allow. Need policy, buy routing you can inspect. Want delegation, buy a blended model with published receipts, and hold it to that standard.
Why does the terminology matter?
Because the blur sells: 'router' pricing pages rarely mention that the routing table is yours to maintain forever. Precise terms make the real question visible: who picks the model, and how often must they re-pick?
Compare any two models
Rates verified 2026-07-28. "Measured task" = our identical dashboard-generation prompt, metered where marked ✓ and list-math otherwise. Verbosity from the Verbosity Index, Edition 1. Data: prices.json.