Where the bitcoin price actually comes from

Somebody asks what bitcoin is worth. You open two tabs and get two answers. Neither site is broken, neither is lying, and the difference is not rounding — it is the whole subject of this page.

The confusion comes from a reasonable assumption imported from equities, where a stock has one last sale because there is a consolidated tape and a rule that produces it. Crypto has no consolidated tape, no regulator mandating one, and no single venue that all the others report into. What it has instead is a few hundred order books, each printing its own trades, and a market of firms selling you different opinions about how to combine them.

How it works

The only primitive in this market is a trade print: one venue's matching engine pairs a buy and a sell and publishes a price, a size and a timestamp. That print is a fact about that venue at that instant. It is not a fact about bitcoin.

Nothing merges the prints. There is no entity whose job it is to hold the union of every venue's tape, and no venue is obliged to tell another one anything. So the moment you want a number that is not "the last trade on this one exchange", you are no longer measuring — you are computing, and two decisions have to be made before a number exists at all: which prints go in, and what function runs over them.

The chain from a print to the number on your screen has four links and money moves along only one of them:

1. The venue generates the print. It owns it.
2. The venue publishes it free, usually on an endpoint that needs no key at all, and reserves every right in it by contract.
3. A vendor licenses the redistribution right, venue by venue, and collects the prints across many of them into one schema, one clock and one symbology.
4. The vendor applies a method and sells you the output.

The licensing half of that chain is its own subject and has its own page here: what exchange API terms actually let you do. This page is about the fourth link — the method, which is the part that decides what number you actually get.

What an aggregation method decides

Four decisions, and every one of them moves the answer. What follows compares two published methods on purpose: CoinGecko's, because it is an aggregator that writes its rules down, and CF Benchmarks', because a regulated benchmark administrator has to.

Which markets count, and in what currency. CoinGecko builds a per-coin ticker set from "the top 600 tickers by volume", and its BTC-USD reference is "the volume-weighted average price (VWAP) of selected BTC/USD, BTC/USDT, BTC/USDC, or BTC/EUR tickers across a selection of reputable exchanges", with the stablecoin pairs converted through an internal USD onramp index and the euro pairs through a third-party FX rate. Read that carefully: a number labelled USD is partly built from markets in which no dollar changed hands. The CME CF Bitcoin Reference Rate does the opposite — a fixed list of exchanges approved by an oversight committee, and exactly one stablecoin on its Accepted Asset List, USD Coin.

How the prints are weighted. A volume-weighted average lets one large trade pull the result toward its price. A volume-weighted median does not: it picks the price at the 50th percentile of volume, so size decides which print is chosen rather than how hard it pulls. The CME CF methodology is explicit that this is the point — the use of medians "greatly reduces" susceptibility to price extremes on one or more constituent exchanges. If you have ever wondered why a benchmark barely moved through a wick that your exchange chart shows clearly, this is why.

Which prints are thrown out. CoinGecko applies a median-absolute-deviation bound to coins with three or more tickers, and for coins with fewer than three treats any price change greater than 100x as an outlier. CF Benchmarks runs two screens: an erroneous-data screen (non-numeric or non-positive price or size, unparseable format, or an execution time more than one minute in the future of the calculation agent's own clock) and a potentially-erroneous-data screen that compares each exchange's volume-weighted median against the median of all the exchanges' medians. For the Bitcoin Reference Rate the tolerance is 5 per cent — past that, every trade from that exchange for that day is discarded and the flag is reported to the oversight committee.

What happens when a venue stops answering. This is the decision that nobody advertises and the one that produces the disagreements people notice. The benchmark side is written down in full: anything not retrievable from a constituent exchange's API by one minute after the calculation window closes is disregarded; a constituent exchange with no retrievable trades is disregarded entirely; a five-minute partition in which nothing traded anywhere is dropped and the denominator decremented accordingly; if nothing at all can be retrieved, a calculation failure is declared; and if no trade occurred on any constituent exchange, a market failure event.

On the aggregator side the published rule is usually about staleness rather than outage. CoinGecko excludes trading pairs "that have been blacklisted for inconsistent data and have not been updated for over 3 hours". That is a clear rule and it is worth reading in the other direction: a market that last printed two hours ago is still in the average. There is nothing wrong with that — it is a deliberate choice for a product whose job is to describe a long tail of thinly traded assets — but it is the choice, not an accident, and it is why "the exchange went quiet" and "the price stopped moving" are different events that look identical from outside.

A reference rate is a different product from a ticker

The word "price" covers two products with different purposes, and the difference is not quality.

A reference rate exists so that a contract can settle against it. Its whole value is that the rule is fixed in advance, published, governed and auditable, which means it must be boring and must be the same number for everybody. The CME CF Bitcoin Reference Rate, which CME's bitcoin futures settle against, is specified in full in a public methodology guide on version 17.4 dated 24 August 2026:

• an observation window of 60 minutes, from 15:00 to 16:00 London time, with an effective time of 16:00 London;
• the window partitioned into twelve equal five-minute intervals;
• for each partition, the volume-weighted median trade price across all constituent exchanges together, not per exchange;
• the rate itself being the equally weighted average of those twelve medians.

That last choice is the tell. The partitions are equally weighted rather than volume-weighted specifically so the rate can be replicated: a desk that must transact at the rate can trade a twelfth of its size in each partition and track it. A number designed to be tradeable is a different artefact from a number designed to be looked at.

The constituent list is published too, with dates. As of version 13.7, also dated 24 August 2026, the Bitcoin Reference Rate runs on seven active exchanges — Bitstamp, Coinbase and Kraken since the launch on 14 November 2016, Gemini from 30 August 2019, LMAX Digital from 3 May 2021, Bullish Exchange from 30 December 2024 and Crypto.com from 31 March 2025 — and the document keeps the suspensions on the record as well: Bitfinex and OKCoin.com (HK) from April 2017, itBit from 27 July 2026.

Seven venues. Against that, an aggregator's marketing page will claim several hundred exchanges, and both numbers are honest, because the two products are answering different questions. One is trying to produce a settleable price for one pair; the other is trying to describe an entire market including assets that trade on two venues in the world.

Governance is the other half of what you are buying, and it is visible for the same reason. CF Benchmarks states in its own documents that it is authorised by the UK Financial Conduct Authority as a registered benchmark administrator, FRN 847100. The constituent-exchange list carries a disclosure that CF Benchmarks sits inside the Payward group, which owns and operates Kraken — one of the constituent exchanges feeding the rate. You are free to think whatever you like about that arrangement; the point is that it is printed on the document, because a benchmark regime requires it to be. A ticker on an aggregator's site carries no equivalent disclosure, because nothing requires one, and its absence is not evidence that there is nothing to disclose.

Two more things a benchmark does that a ticker does not: it publishes a dated version history, so you can read every revision the methodology has had since version 1.0 of 10 November 2016 and what each one changed; and it defines, in advance, what happens when it cannot be calculated at all.

What it costs

The prints themselves cost nothing. Coinbase Exchange, Kraken and Bitstamp all answer a BTC-USD ticker request over plain HTTPS with no account and no key, and so do most of the venues you have heard of.

What is not free is the right to publish a number derived from them, and — specifically — the right to build a benchmark out of them. Coinbase's Market Data Terms of Use, last updated 7 August 2026, prohibit using its market data, absent prior express written consent, to create "indexes, fixings, or other benchmarks", "generic or fair value prices", or valuations of a financial product, and separately prohibit using the data "as a benchmark", a phrase the document then defines at length. An index administrator that wants Coinbase prints in a settlement rate is buying a licence, not calling an endpoint. Neither the licence nor its price is published anywhere.

So the market splits into three price points, and the middle one is where most readers actually live:

• Zero. A single venue's own API, for that venue's own tape — authoritative for that venue and licensed for your own use only. Or an aggregator's free tier, which typically costs nothing, caps you in credits per month and calls per minute, requires attribution, and forbids redistribution. CoinGecko and CoinMarketCap both run one, and Coin Metrics publishes a community reference rate that needs no key at all.
• Published subscription pricing, in the tens to hundreds of dollars a month, for licensed aggregated data at usable rate limits. The cards in market data APIs carry the current figures per vendor.
• Quote only. Index licences, tick archives and anything that settles money. Kaiko and Coin Metrics above their free tiers both sell this way, and no vendor in this part of the market publishes a rate card. Budget for a negotiation rather than a checkout.

What you can do about it

First, work out which of three questions you are actually asking. Almost every bad choice in this area is a category error rather than a pricing mistake.

1. What did it trade at somewhere I can execute? Use that venue's own API. It is free, it is the only authoritative source for its own tape, and no aggregate will ever match your fills better.
2. What is the broad market level, for a dashboard, a portfolio total or a chart? An aggregator ticker is the right product and the free tier is usually enough. Accept that it will disagree with your exchange by a few basis points and do not build anything that cares.
3. What number settles a contract, values a fund, triggers a liquidation or goes into an audited figure? A reference rate with a published methodology, a named administrator and a licence that permits that use. Nothing else will survive being asked where the number came from.

Never let one system mix them. A liquidation engine reading an aggregator ticker and an execution path reading a venue book is a design that works until a venue goes quiet.

Before you commit to a provider, get four answers in writing. Which venues are in, and how you are told when that changes. How they are weighted. What the outlier rule is. And what the feed does when a venue stops responding — drop and reweight, hold the last print, or keep publishing. The fourth one is rarely on the pricing page and it is the one that will cost you.

Check that every value carries its own timestamp, and use it. CoinGecko's simple-price endpoint returns last_updated_at, CoinMarketCap returns last_updated per quote, an exchange ticker returns the time of the print. If your source gives you a number with no per-value timestamp, you cannot distinguish a live price from a held one, and that is a property of the feed, not of the market.

Measure the disagreement yourself before you care about it. It takes one minute and no keys:

curl -s "https://api.exchange.coinbase.com/products/BTC-USD/ticker"
curl -s "https://api.kraken.com/0/public/Ticker?pair=XBTUSD"
curl -s "https://www.bitstamp.net/api/v2/ticker/btcusd/"
curl -s "https://api.coingecko.com/api/v3/simple/price?ids=bitcoin&vs_currencies=usd&include_last_updated_at=true"
curl -s "https://community-api.coinmetrics.io/v4/timeseries/asset-metrics?assets=btc&metrics=ReferenceRateUSD&frequency=1s"

Running exactly that here at 14:13 UTC on 19 September 2026 — our own reading, not a citation, and reproducible only in shape — gave three single-venue last trades within 10 USD of each other (81,356.83 on Coinbase Exchange, 81,355.00 on Kraken, 81,347.54 on Bitstamp), a second-by-second reference rate at 81,356.82, and two aggregator tickers at 81,385 and 81,387.65. Total spread about 40 USD on a price of 81,000, or roughly five basis points, on an ordinary afternoon with nothing happening.

Two things in that are worth keeping. The aggregators sat above all three venue prints; nothing there proves why, but the mechanism that would produce it is in CoinGecko's published method, which feeds its BTC index from USDT and USDC pairs as well as dollar pairs. And five basis points is the quiet-day number. The gap widens exactly when it matters — during the minutes when venues diverge, one of them stops responding, or a large trade lands on a thin book — which is the case your design has to survive, and the one you cannot see by checking on a Tuesday afternoon.

If you are the one computing a number, read the terms before you write the code. The benchmark-creation clause quoted above is the one that catches people, because it is not a redistribution restriction and it fires even on work you never publish.