Gateway, aggregator, or direct: get the taxonomy right first.

Updated 2026-07-16

The words gateway and aggregator get used interchangeably in marketing, and the confusion costs teams real architecture mistakes. The taxonomy is actually clean once you ask two questions: who runs the routing layer, and who holds the vendor accounts. This page sorts the three operating models and gives you the decision path.

Quick answer: two questions sort every product.

Question one: who runs the routing layer, the software that accepts your request and forwards it to a model vendor? Either nobody (you call vendors directly), you (a router deployed in your infrastructure), or a third party (a hosted endpoint). Question two: who holds the vendor accounts and pays the vendors? Either you hold a billing relationship per vendor, or the routing operator holds them and resells capacity to you as one bill. Direct integration answers nobody/you: no router, your accounts. A self-hosted gateway answers you/you: your router, your accounts, list prices. A hosted aggregator answers them/them: their router, their vendor accounts, one key and one bill for you. There is a fourth cell, a hosted router where you bring your own vendor keys, which some cloud gateways offer; it behaves like a self-hosted gateway operationally, minus the self-hosting. Every product in this market sits in one of those cells, whatever its landing page calls itself. APIsRouter, for the record, is a hosted aggregator: one OpenAI-compatible endpoint, one prepaid balance, the vendor relationships held on our side.

How each model works, mechanically.

Direct: your service imports each vendor's SDK or calls its HTTP API, authenticated with that vendor's key. Requests travel straight to the vendor. You own retries, rate-limit handling, failover, usage rollups, and one billing relationship per vendor. Nothing sits between you and the model, and nothing helps you either. Self-hosted gateway: you deploy a router, open-source projects like LiteLLM and one-api are the canonical examples, inside your own infrastructure. Your application calls the router on one OpenAI-compatible interface; the router holds your per-vendor keys and forwards each request by model id, normalizing shapes where vendors differ. You gain the unified interface, central key custody, and one place for retries and metering; you also acquire a service to operate, upgrade, and monitor, and your token prices remain each vendor's list. Hosted aggregator: your application calls a third-party endpoint with one key. The aggregator's router forwards to vendors under its own accounts and bills you per token or per image from one balance. You gain everything the self-hosted router gives, with zero operations, no per-vendor sign-ups, and often unit prices below vendor list, since aggregators can pool capacity. You accept a third party in the request path, both for availability and for data transit, which is the honest cost of the lane.

The taxonomy in one table.

Same two questions, laid out with their consequences. For the deep two-way comparison of direct versus gateway-shaped architecture, including a worked month of production traffic priced both ways, see the companion guide linked below; this table stays at the taxonomy level.

The two bolded questions, who routes and who bills, generate every row.
PropertyDirectSelf-hosted gatewayHosted aggregator
Routing layer run byNobodyYouThe provider
Vendor accounts held byYou, one per vendorYou, one per vendorThe provider
Your integrationOne per vendorOne, against your routerOne, against their endpoint
Keys you manageMany vendor keysMany vendor keys, centralizedOne key
Bills you receiveOne per vendorOne per vendorOne
Token pricingVendor listVendor listProvider catalog, often below list
Ops burdenRetry/failover code per vendorA service to runNone
Data pathYou to vendorYour infra to vendorTransits the provider
New model accessNew vendor accountNew key plus router configAlready in the catalog

Where the marketing muddies it.

The practical defense against terminology soup is to ask the two taxonomy questions of any product you evaluate. Five minutes with the pricing page answers both: if there is a per-token price sheet, vendor accounts are theirs; if there is a deployment guide, the router is yours.

  • Gateway is used for both the self-hosted routers and the hosted resellers; when a product page says gateway, check whether you bring vendor keys (router) or top up a balance (aggregator). The billing model is the tell.
  • Aggregator and unified API usually mean the reseller cell, but some unified APIs are bring-your-own-key routers hosted in someone's cloud: router semantics, hosted convenience.
  • Proxy usually describes a thin pass-through in front of one vendor, a special case of the router with a catalog of one.
  • Router is overloaded twice: infrastructure routing (this page) and model routing, choosing which model answers a prompt. A product can do either or both; the words do not distinguish, the feature list does.
  • OpenAI-compatible is orthogonal to all of it: every cell can expose the same /v1 interface, which is precisely why your application code should not care which cell serves it.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

ModelOfficial PriceOur Price
Claude Sonnet 4.6$3.00 / $15.00 per M$2.40 / $12.00 per M
GPT-5.5$5.00 / $30.00 per M$4.00 / $24.00 per M
Gemini 3.1 Pro Preview$2.00 / $12.00 per M$1.60 / $9.60 per M
DeepSeek V4 Pro$0.43 / $0.87 per M$0.39 / $0.78 per M

The decision path.

One provider, provider-exclusive features, or compliance that forbids intermediaries: go direct. This is a genuine lane, not a fallback; a single-vendor product gains nothing from a router it does not need. Multiple model families, strict data-path control, and an infrastructure team that will own another service: self-hosted gateway. You get unified routing with your traffic never leaving your perimeter except to vendors, and you pay for it in operations and in list-price tokens. Multiple model families and a preference for shipping over operating: hosted aggregator. One integration, one key, one bill, models from every major family, and unit pricing that is usually the best of the three lanes. Accept the third party in the path deliberately, read the provider's data handling terms, and keep the escape hatch below. The escape hatch is the same for every lane, and it is the most important sentence on this page: code against the OpenAI-compatible surface with base URL and model id as configuration. Do that, and the three cells become deployment choices you can revisit per environment, staging on an aggregator, production direct, or the reverse, rather than architectural commitments. Teams migrate between cells in an afternoon when the interface stayed portable, and rewrite for a quarter when it did not.

Mixing lanes is normal, not indecision.

Mature stacks usually occupy two cells at once, on purpose. A common shape: the aggregator serves development, evaluation, and the long tail of models, because one key against a broad catalog is unbeatable for iteration speed, while one anchor vendor stays direct for a compliance-scoped workload or a provider-exclusive feature. Another: a self-hosted router fronts everything on-perimeter, with an aggregator configured as one of its upstreams for the families the team has no direct account with. Both shapes fall out of the same portable-interface discipline, and both keep the aggregator's strongest property in play: when a new model ships from any vendor, it appears in the catalog behind the key you already have, and trying it is a one-string change. That, more than any single month's unit price, is what the reseller cell is actually selling: option value on a fast-moving model market.

FAQ

What is the difference between an AI gateway and an aggregator?

Who holds the vendor accounts. A gateway (in the precise sense) is a router, usually self-hosted, that forwards requests using your per-vendor keys at vendor list prices. An aggregator hosts the router and the vendor relationships, reselling capacity as one key and one bill, typically below list. Marketing uses the words loosely; the billing model reveals the truth.

Is APIsRouter a gateway or an aggregator?

A hosted aggregator by this taxonomy: one OpenAI-compatible endpoint, the vendor accounts held on our side, one prepaid balance and key for the whole catalog. We use the word gateway informally on the site, which is exactly the loose usage this page equips you to see through.

When should I stay on direct vendor APIs?

Single-provider products, dependencies on provider-exclusive beta features, and compliance regimes that forbid an intermediary in the data path. In those cases the router adds a hop without adding value. The companion deep dive on gateway versus direct covers this trade in detail with cost math.

What are examples of self-hosted gateways?

LiteLLM and one-api are the widely deployed open-source options: you run the router, bring per-vendor keys, and get a unified OpenAI-compatible interface inside your own infrastructure at vendor list prices. Choose this cell when data-path control justifies operating another service.

Do aggregators actually price below vendor list?

Often, and it is structural rather than promotional: pooled capacity across many customers. On APIsRouter, catalog prices per model are published on the pricing page next to official list for comparison, so the claim is checkable per id rather than taken on faith.

Can I use more than one of these at once?

Yes, and most mature stacks do: an aggregator for breadth and iteration, direct for an anchor vendor or compliance-scoped traffic, or a self-hosted router with an aggregator as one upstream. Keeping base URL and model id in configuration is what makes the mix painless to operate and change.