Paid AI API Routers

The category is defined by the billing contract, not by the absence of free access. The 100 entries here are endpoints where per-use billing — tokens in, tokens out, charged at published per-1M rates — is the primary commercial shape: the labs' own developer APIs, model endpoints hosted on cloud marketplaces, and inference clouds serving open models on metered usage. An entry can sit in this category and still offer a free on-ramp — 50 do, with a quantified allowance. The category records the paid contract because that is what a production workload will eventually sign.

Two sibling categories hold the other shapes, and the boundary is worth internalizing. Official free tier entries are funnels: free now, by design, to pull you into the same provider's paid program. Commercial aggregator entries sell a single key across many upstreams, so their price is a routing layer over someone else's rate card. Paid API is the direct contract: one provider, one rate table, one invoice — and the normalization on each entry exists to make those rate tables comparable across providers that price different model families.

Each entry shows two numbers: the cheapest tracked input rate and the cheapest tracked output rate, per 1M tokens. Output almost always costs more than input, and the spread is one of the real decision variables — a workload that resends long conversation histories every turn is input-heavy; a generator is output-heavy, and provider rankings can flip between the two. One concrete published example from the sources: Grok 4.7 at $2 per 1M input and $6 per 1M output (xAI docs, updated 2026-09-21). Your own input/output ratio is what moves your effective rate across the spread.

The entry's number is the minimum across tracked models, not the rate you will pay: it exists so rows can be compared at a glance, the way a flight search shows the cheapest fare. The full per-model table lives at the provider — each entry links its pricing source — and a workload that needs the frontier model may pay several multiples of the minimum. Where a provider bills outside the token unit (per request, per generation, per minute) the dataset normalizes at the provider's published rates and flags the conversion ; a provider with no fixed rate shows no price rather than a sentinel, because a fake number would corrupt the comparison.

Published standard rates are the ceiling of the discount structure, not the only option, where the provider documents alternatives. Batch: both Mistral and AWS Bedrock publish a 50% discount for asynchronous, higher-latency workloads — the classic fit is anything that can wait hours: evaluation runs, data backfills, offline generation. Cached input: several providers price a repeated prompt prefix below the standard input rate when the prefix is cacheable — long system prompts and few-shot blocks are the beneficiary. These are documented discounts, and the entry notes carry them where tracked.

Service tiers are the second axis: DeepInfra, for example, documents standard, priority and flex tiers where the price trades against scheduling priority and burst headroom. Choosing a tier is a reliability decision wearing a price label — the same model, a different queue. The dataset records the tier structure on the entry where the provider publishes it; it does not benchmark the queueing itself, so treat tier comparisons as the provider's claim until your own workload has measured it.

Free credits are a one-time grant: a fixed amount, often a fixed window, and a budget that stops scaling the moment it runs out. Per-use pricing is linear: cost grows with tokens, but the rate is stable and the ceiling is your budget, not the grant. The breakeven is a volume question. A workload that sends a few million output tokens a month at a mid-range rate spends a fraction of what a typical credit grant is worth — free allowances win by default. A workload in the hundreds of millions of tokens a month is past any grant: unit economics, batch discounts and committed-use options (the enterprise category) become the whole conversation, and the free tier is a prototype budget, not a budget.

There is a middle band where the honest answer is "run both": the free or discounted on-ramp for development, evaluation and low-traffic paths, the per-use key for production, with the split a configuration change. Many of the 100 entries support exactly that — a free allowance and a metered path from the same provider, so the model behavior you test is the model behavior you ship.

Rate tables change without ceremony: models are re-priced, new tiers appear, discounts widen or close. Two mechanisms keep this page honest. First, the entries: each carries a price-change history (when the tracked minimum moved, from what to what) and a last-verified date — a drift between the two is a signal to re-check the provider's table before you budget on it. Second, the sources below: they are the providers' own pricing pages, fetched and checked on the verification date, and they outrank everything on this page the moment they disagree with it. The xAI model docs in the sources carry an "updated" stamp of 2026-09-21 — an example of exactly the drift this section exists to track.

Four checks, in order. One, capability first, price second: identify the cheapest model in the entry that meets your task bar on your own prompts — not the entry's minimum, which may belong to a model your task rejects. Two, volume shape: estimate monthly input and output tokens separately (they are priced separately), then compute the effective rate at your ratio; if any share of the volume is latency-tolerant, price the batch option into the comparison. Three, discount structure: does the provider document cached-input, batch or tier discounts that fit your traffic — a 50% batch discount is worth more than a $0.10 spread on standard rates for a backfill-heavy workload. Four, contract shape: when volume, data terms or compliance start to matter, the same model family often has an enterprise offering — check the enterprise category before you start negotiating.

What this page does not decide for you: model quality for your task (only your evaluation prompts can), queue behavior under your peak (only your traffic can), and terms (only the provider's current page can). It does decide the smaller questions — which providers track a free on-ramp, what the published rates were on the verification date, and which entries have moved their prices since.