Crypto Exchange API Latency: Measured, Not Guessed

We hit 9 crypto exchanges' public market-data APIs with a persistent connection and time the round trips, from a fixed US East server. Lower is better for bots, market-making, and anything latency-sensitive.

Last measured 2026-08-21 06:07 UTC · measured from a fixed US East server (New York) · 50 samples/venue ·raw JSON · CC BY 4.0 (cite FillBench)

Also probed and blocked from US datacenters: Binance.com (HTTP 451), Bybit (HTTP 403). Worth knowing before pointing a bot at them from US infrastructure.

How p95 latency has moved over time

• Coinbase9.95 ms
• Binance.US11.95 ms
• Bitstamp32.51 ms
• Kraken52.44 ms
• Gemini124.07 ms
• Crypto.com188.27 ms

Put this on your site (free)

Building a bot or writing about exchanges? Drop this live leaderboard on your page. It re-measures on its own, so it stays current with zero upkeep, and it always shows the latest run straight from this benchmark. One line of HTML, no script, no signup.

<iframe src="https://fillbench.com/embed/latency" width="460" height="340"
        style="border:0;max-width:100%" loading="lazy"
        title="Live crypto exchange API latency by FillBench"></iframe>

How to read this

Connect is the one-time DNS + TCP + TLS handshake cost. A running bot pays it once at startup and then reuses the connection, so it barely affects steady-state speed. It's the median of several handshake samples, so it's approximate but not just noise; a bot pays it once, so we do not rank on it.p50 is the typical request. p95 / p99 are the slow tail, the requests that actually cost you during volatile markets, which is why we rank on p95: a venue with a great median but an ugly tail will still burn you. The US column flags whether a venue serves US retail customers (verified 2026-07): a fast API you cannot legally use is not much help, so we show it rather than quietly drop the venue. The "fastest" verdict above is scoped to venues a US trader can actually use.

Where we measure from: every run executes on a dedicated server in US East (New York), the same box every time. That fixed vantage point is the point: it keeps the comparison between exchanges stable from run to run instead of drifting with whatever cloud region happened to host the test. Your own latency still depends on where your bot runs, which is exactly why colocation and VPS choice matter. A bot in the same region as the exchange will see numbers close to ours; one on the other side of the world will not. Read this as a fair relative ranking measured from one consistent US vantage point.

Trading on these exchanges

Latency is one input. Fees, available pairs, API rate limits, and US eligibility matter too. We measure the real cost to trade, base-tier taker fee plus live spread and slippage, in our exchange fee benchmark. If you're building or running a bot, pairing a low-latency exchange with a nearby VPS is the single biggest controllable win. Many traders fire their entries from TradingView alerts, and the plan tier you need scales with how many alerts your strategy runs.

A different kind of speed, worth not confusing with this one: everything above is exchange API latency, how fast a venue's server answers a request. A separate question, once your bot becomes a paying agent, is how fast the payment itself finalizes on the blockchain, which is an entirely different measurement. We benchmark that per chain in our x402 blockchain settlement benchmark (Algorand finalizes a payment in about 5.2 seconds, measured on-chain).

Common questions