Commercial Aggregator AI API Routers

A commercial aggregator is a company that fronts inference from multiple upstream providers — the labs, the open-model clouds, the specialty endpoints — behind a single API. You authenticate once, address models by name (usually an org/model slug), and pay one bill. OpenRouter is the archetype the whole category is measured against: hundreds of models through one endpoint, with published routing policy (cost-aware provider selection with automatic fallback when an upstream returns errors or rate-limits) documented in its own API reference. Hugging Face's Inference Providers run the same shape across eighteen partner providers and state "no extra markup on provider rates" — the margin question answered explicitly, which is rarer than it should be.

The category exists because the direct-to-lab route has real friction: N consoles, N keys, N billing relationships, N rate-limit regimes, N outages. An aggregator collapses that to one of each — and adds things labs don't offer, like failover across providers for the same model and a single ledger to reconcile. What it costs is the difference between the lab's price and the aggregator's price, plus a third party in your data path. Both costs are checkable per entry, which is what this page is for.

Each entry carries per-1M-token prices (input and output, the cheapest tracked model for each), the free-credit value where the aggregator offers a trial or recurring allowance, and the verification date for both. Two things deserve attention. First, the spread: compare the aggregator's rate for a model against the lab's own published rate for that model (the lab entries in this directory carry them) — the difference is the effective markup, and it varies by model and by aggregator, so "aggregators are 5% more expensive" is a fiction; some are at parity, some are meaningfully above, and a few price below lab list to win traffic. Second, the free side: aggregators compete for developers the same way labs do, so several carry signup credits or monthly allowances — the credit column on this page is exactly that layer, ranked.

There is a subtler cost the table doesn't carry: batching and caching discounts. Several labs (OpenAI's pricing page is dense with cached-input rates; AWS Bedrock publishes 50% batch pricing) discount heavy workloads in ways an aggregator may or may not pass through. If your workload is batch-shaped, check the aggregator's terms on discount pass-through before you trust the headline per-1M number.

When you call a model through an aggregator, your prompt transits the aggregator's infrastructure on the way to the upstream. For most workloads this is unremarkable — the aggregator is a vendor, like any API host. For sensitive workloads it's the central terms question: what does the aggregator retain, log, train on or route where? Established aggregators publish this (OpenRouter documents request routing and retention behavior; Hugging Face's provider documentation names the partner chain per model), and the risk rationale on an entry points at the documentation where it exists. Where it doesn't exist, the rating says "unclear" rather than assuming clean — read that as the signal it is.

We track catalog size (tracked models per entry), published per-1M rates with dates, free-credit allowances normalized to USD ( where our assumption is inside the math), operator risk evidence and endpoint availability via probes. We do not benchmark aggregator latency against direct lab calls (the gap varies by model, provider routing and time of day), and we do not audit the data path — we point at the documentation that describes it. The compare tool lets you put two aggregators, or an aggregator and a direct lab entry, side by side on the fields we do track.

Aggregators win on the integration axis: you need many model families under one key, you want failover across providers without writing it yourself, you want one bill for finance to reconcile, or you're early in a project and don't want to re-point your code every time the model choice changes. OpenRouter's documented routing — pick the least expensive healthy provider, fall back on 5xx and rate limits — is the canonical example of value you literally cannot get from a single lab. They win on the catalog axis too: long-tail open models that only exist on one specialty host become one slug away.

Direct beats aggregator when the math or the terms demand it: your workload is heavy enough that the markup is a real line item (compare per-1M rates for your specific model), you need a discount regime the aggregator doesn't pass through (batch, cached input), your data terms require a direct relationship (some enterprises will only sign with the lab), or you're at volumes where the lab's sales team writes custom pricing that makes any aggregator irrelevant. The decision is per-workload, not per-company — which is why the entry-level rates and terms links matter more than the category story.