Crypto Market Data APIs

What this category is

APIs you call from your own code to get prices, order books, trades and the reference data that ties them together. The buyer writes software — that is the line against the rest of the catalogue. A dashboard you log into to read the same numbers is a different purchase, and a product whose value is a chart on a screen belongs under charting or on-chain analytics.

What is being sold here is not really data — the exchanges publish most of it for free — but normalisation, uptime and history. One symbol scheme across two hundred venues, a feed that keeps producing when a venue drops out, and a tick archive going back further than the venue itself keeps. Those three are what the price is for, and the three hardest things to check from a pricing page.

Five jobs share this page, and they barely compete

Sixteen products call this category home, and each card carries a job facet. Reading the category by that facet is more useful than reading it by vendor size, because a buyer almost never shortlists across two of them.

Live prices for an application is the largest group and the one every comparison article is about — CoinGecko, CoinMarketCap and CoinAPI on the aggregator side, the venues' own endpoints (Binance, Bybit, OKX, Coinbase) on the other. Four of the sixteen cards are an exchange publishing its own tape: free, fast, and blind to every venue but one.

History, sold by the byte. Tardis.dev, Kaiko, CoinDesk Data and Massive are bought for what happened rather than what is happening. This is where the price is highest and the free tier thinnest, because the archive is the asset.

Derivatives structure. Amberdata, Laevitas and CoinGlass publish funding, open interest, liquidations and options surfaces that the spot aggregators largely do not normalise at all.

Benchmarks, and everything around them. Coin Metrics is filed under reference data — rates and indexes with a published methodology, for people who need to point at the rules rather than the number — and Messari under news, here for research and asset profiles as much as prices.

The use of that split is negative. If you need funding rates, most of these cards will not sell them at any price. Work out which of the five you are in before reading a pricing page.

Three words that mean nothing until you pin them down

"Real time" covers everything from a websocket tick pushed on trade to a REST endpoint served from a fifteen-second cache. Both print as "real-time data", and fifteen of the sixteen cards here record their latency as real-time, which shows how little the word discriminates. The one that does not is Tardis.dev, and that is the honest case rather than the weak one: an archive delivered as files is not a feed. What matters is the update interval on the tier you are buying, and whether websockets are on that tier at all — several providers gate them behind a paid plan while advertising real-time data on the free one, which is true of the product and false of the tier.

"Historical" covers everything from two years of daily candles to full order-book snapshots from 2013. The useful question is not how far back but at what granularity how far back — daily OHLCV for ten years and tick data for three months sit behind the same word on the same page. A candle is an aggregation with the publisher's choices baked in, so two vendors' one-minute bars for the same pair are routinely two different series before history enters the conversation. Ask for the oldest timestamp you need, at the resolution you need it, before you pay.

"Supported exchanges" is a count, and counts include venues polled once a minute next to venues with a direct feed behind them, plus venues that have stopped trading — removing a name lowers a number on a marketing page. A count of 200+ tells you less than one name and its latency.

The question that decides it, and you can answer it today

Does the data leave your organisation?

If the answer is no — an internal dashboard, a model you train, a bot trading your own book — the cheapest correct answer is usually an exchange's own API, and the four venue cards exist so you can check that before paying anyone. You give up cross-venue normalisation and keep the money.

If the answer is yes — a chart on your product, a price in your app, a number in a report you send a client — then redistribution decides, and it is a licence question rather than a technical one. It is rarely on the pricing page, routinely in the API terms, and it rules out the free route for several major venues. The paid redistributor layer exists precisely because the free endpoints do not carry the right, which what exchange API terms actually let you do works through in full.

The follow-up is whether you need DEX prints. Nine of the sixteen cards report decentralised venues alongside centralised ones; the seven that do not are the four exchange APIs plus Massive, Tardis.dev and Laevitas. A pair that only trades on-chain is invisible to those seven, and no plan upgrade fixes it.

What the pricing actually looks like from inside

This is the most quoted-on-request category in the catalogue — three quarters of the cards carry a tier with no number on it, against none at all in the backtesting libraries. The per-category counts and what the gradient means are on tools you cannot buy without a sales call and why this market won't quote you a price.

Two things about that shape are specific to this category.

The published prices are two clusters with a hole between them. Read the cheapest paid month on each card and five sit at or below 79 dollars — CoinGecko, CoinMarketCap, CoinGlass, Massive and CoinAPI — while three sit at or above 350 — Tardis.dev, Laevitas and Amberdata. Nothing is priced in between. That gap is the boundary between self-serve data and licensed data, and crossing it is a procurement decision rather than an upgrade.

Only four cards carry a number on every tier they publish: Binance, Bybit, OKX and Tardis.dev — three exchanges giving away their own tape, one vendor selling files. Everybody else keeps a door marked "contact us", and on CoinDesk Data and Kaiko that is every door.

What to check before you commit

Whether the rate limit is per key or per IP, and what unit it counts. A per-IP limit that looked generous becomes the binding constraint the moment the service runs on more than one container. And credits, requests and weights are not the same unit — a plan advertised at ten thousand calls may cost a hundred weighted units per call on the endpoint you need. Normalise against your own query before comparing prices.

How bulk data comes out. Every card here returns JSON and most also offer CSV, but a year of ticks is not a JSON response. Two list Parquet — Amberdata and Coin Metrics — and both sit at the quoted end. If the plan is to load history into a dataframe, the delivery format costs more than the price does.

What happens during an outage on the venue underneath. An aggregate price is a method — some providers drop the failed venue and reweight, some hold the last trade, some keep publishing a stale number without saying so. That is where two "accurate" providers disagree by a percent and neither is lying, and it is the subject of where the bitcoin price actually comes from.

What is not here

The exchanges themselves. A venue's public data endpoint is a product on its own terms and gets a card; the venue is not reviewed, rated or recommended anywhere on the site.

Chain data sold as chain data. Wallet balances, holder distributions, protocol revenue and labelled flows are on-chain analytics, even when a DEX trade is also a market event. RPC endpoints and indexing backends are plumbing, and out of scope entirely.

Anything you look at rather than call. A screen showing funding across forty venues is a screening product; a screen you can also trade from is a terminal. Several vendors here sell both and get two cards, priced separately, because they are two purchases — which is why a vendor's name in the search box returns more than one result.

Products that have already gone. CryptoCompare, whose terms changed underneath its users, and Cryptowatch, which shut down, have pages of their own. Read one before building against a free endpoint that feels permanent.