Credit, request, weight, symbol: why API quotas don't compare

Every data API in this catalogue publishes a quota, and almost every one of them counts something different. The ceilings themselves — per key or per IP, per second or per month — are covered under rate limit. This page is about the unit underneath the ceiling: what one credit, one call or one point of weight actually buys, why the same job produces a different number on every provider, and how to turn those numbers into something you can compare.

How it works

Strip the branding off and there are five things a vendor can count, plus a sixth arrangement that is not a meter at all.

The request. One call, one unit, whatever comes back. CoinGecko deducts one credit from your monthly balance for each successful call, and Kaiko allows 6,000 requests a minute per API key on its standard REST API. Planning against a request meter is simple: count calls.

Weight. Still one call, but the vendor sets its price per endpoint and sometimes per parameter. The Binance API charges 2 for a candle request and 250 for a full-depth order book, out of 6,000 a minute per IP. Nansen does the same with credits: most profiler and token endpoints cost 1, Smart Money endpoints 5, an address lookup for common labels 100, the premium-labels endpoint 500, and an expert-mode agent question 750. A monthly Nansen balance means little until you know which endpoints will spend it.

Data volume. The unit is how much comes back rather than the call that fetched it. CoinMarketCap charges one credit per 100 OHLCV data points, rounded up. CoinAPI charges one credit per call, except that a call using the limit parameter costs one credit per 100 data points returned, rounded up, with 1 as the floor. CryptoQuant says only that the more rows come back, the more a call costs, and that different data types cost different amounts. It publishes no rate card.

The asset. Glassnode charges per asset per request: 1 credit for Bitcoin, 2 for any other asset. A bulk request is charged per asset requested, and the number of exchanges in it does not change the price. The same series costs twice as much for Ether as for Bitcoin.

Compute. Dune charges credits for the CPU, data scanned and engine time a query uses, and states that it publishes no per-query formula: "You'll see costs after execution." Reading the result is billed separately, in data points. So one query costs two things, and only the second one can be worked out in advance.

The entitlement. Here the vendor sells the right to ask for a slice of data. The meter is still there, but it is not what decides the price. Amberdata sells self-serve subscriptions one exchange and one market at a time. Its own documentation gives the example of a subscription for Coinbase spot (named gdax in the API): a tickers call that names the exchange succeeds, and the same call without the exchange parameter is rejected. So is a call naming Kraken, and so is a futures call naming the same venue. Kaiko builds each quote from the number of assets or instruments, the data type, the granularity, whether you need history or live data, and usage. On both, the symbol is part of the contract, not an argument you can change.

Scope shows up on the exchange side too. OKX counts its funding-rate and open-interest endpoints against your IP and the instrument together. Polling fifty contracts gives you fifty separate small budgets, not one large one.

One job, seven meters

Take one small job and price it everywhere: a year of daily prices for 20 assets, fetched once. Each row below uses the cheapest documented way to do it and the vendor's own published cost for that endpoint.

Read down the last column and nothing lines up: 40, 20, 80, 80, 39, "unknown plus about 22", and a subscription. None of these is a cheap or an expensive result, and the order changes as soon as the job does. Ask for hourly bars instead of daily ones and the data-volume meters grow 24-fold, because they charge for points. Binance barely moves, because a request still returns up to 1,000 bars for weight 2. Glassnode's credit count does not change at all, because it charges per asset per request rather than per point, although finer resolutions sit on dearer plans. Swap the 20 majors for 20 small tokens and Glassnode nearly doubles, because only Bitcoin costs 1.

The Dune row has a second problem. Dune's pricing FAQ prices API exports at 1,000 data points per credit on Analyst and 5,000 on Plus. Its API billing page prices the same exports at 10 credits per megabyte on Analyst and 2 on Plus, and does not say how a result smaller than a megabyte is rounded. Both pages were live when this guide was checked. Until Dune reconciles them, measure the read cost on your own account rather than assuming either figure.

The Amberdata row fails the comparison in a different way. The subscription covers one venue. If the 20 assets need two exchanges, that is two subscriptions. The job's cost moves with where the data comes from, not with how much of it you take.

What it costs

A unit is also defined by when it comes back and by what happens when it runs out, and those rules differ as much as the units do.

How long a unit lives. Binance weight refills every minute and cannot be saved. CoinAPI's included credits are daily. CoinGecko, CoinMarketCap, CryptoQuant and Dune grant credits monthly. Neither CryptoQuant nor Dune carries unused credits into the next month. Nansen's Pro plan tops the included balance back up to 2,000 on the first of each month, but only if it has fallen below 2,000. That makes it a floor rather than a grant. Nansen's free plan gets 100 credits once and then a daily top-up back to 10.

What happens at the limit. On Binance you get a 429, and an IP that keeps calling after that is banned automatically for two minutes to three days, with the ban growing for repeat offenders. CryptoQuant returns a 402 payment_required once both its plan credits and any purchased credits are used up. CoinAPI does not stop you: usage beyond the plan quota is billed on a daily tiered scale, from $5.26 per 1,000 credits for the first 1,000 of the day down to $0.03 per 1,000 above 25 million, and the tiers reset every day. Dune bills extra credits at $1.875 per 100 on monthly Analyst and $1.596 on monthly Plus, up to a limit you set yourself. It has no free fallback: since 10 September 2026 a free Dune account is view-only after a 14-day trial of 2,500 credits, and free accounts cannot buy extra credits.

What a top-up buys. CryptoQuant sells 100,000 credits for $5, only on its Advanced and Professional plans, and the balance stays valid for 12 months from the last purchase. Nansen's purchased credits expire one year after purchase and are spent only after the included ones. Binance's public endpoints have no price list, so there is no top-up to buy.

So "credits" can mean a monthly balance that never rolls over, a floor that refills on the first of the month, or a daily quota with metered overage behind it. Before any of those figures can be compared, each has to be multiplied out to your job and to the period you are billed in.

What you can do about it

Write the job down as calls before you open a pricing page. List the endpoints, the parameters and how often each runs: symbols, resolution, history depth, refresh interval. Then price each line with the vendor's published cost for that endpoint, not the headline allowance. That is the "one job, seven meters" table above, done for your job, and it takes an afternoon. It will also catch the parameter that multiplies the cost. On Binance, leaving the symbol off a 24-hour ticker request raises its weight from 2 to 80. On CoinMarketCap, each convert currency beyond the first adds a credit.

Put the job in the unit you are billed in. If the plan is monthly, multiply the per-minute or per-day figure out to a month before comparing it with a monthly allowance. If it is metered, price the overage rather than the allowance, because the overage is what a busy month costs.

Run it on a trial, and read the meter the vendor exposes. Binance returns the running total in an X-MBX-USED-WEIGHT-1M header on every response. Dune reports execution_cost_credits for each execution on its status endpoint, which is the only way to learn what a query cost. CryptoQuant shows each call's deduction in its usage history. On a paid plan, CoinGecko's /key endpoint reports monthly and remaining credits. CoinAPI warns that its X-RateLimit-Remaining and X-RateLimit-Reset headers are left off some responses on purpose, so a client that budgets from them has to cope when they are missing.

Cap the compute meters before you automate them. On Dune, set the per-execution cost cap and the monthly extra-credit limit before a scheduled query or an agent can run on its own. Nothing else bounds a meter whose price is published only after the query runs.

For entitlement pricing, list the scope before asking for a quote. For Amberdata or Kaiko, bring exchanges, markets and instruments by name, plus granularity and history depth. Then check the one behaviour that breaks code you already have. On an Amberdata exchange subscription, any call that does not name the subscribed exchange is rejected, even if the same call worked on a trial key with access to every endpoint.

Ask five questions that the pricing page rarely answers. What does one credit buy on the endpoint I will call most? Is it charged per call, per row, per asset or per unit of compute? Do failed calls cost anything? Does unused allowance roll over? What happens at zero: a 429, a 402, an overage bill or a ban? Two vendors that answer all five can be compared. Two that quote only a monthly number cannot.

For what the same providers let you do with the data once you have it, see what exchange API terms actually let you do. The price and market-data APIs among them are listed under market data APIs.