Arbitrage bots

The spread after costs, on one chart

Bot	Type	Price	Arbitrage	Backtest	Paper	Exchanges	US
Hummingbot	Self-hosted	free	yes	yes	yes	50	US
Gunbot	Self-hosted	$29/mo	yes	yes	yes	20	US
Bitsgap	Cloud platform	$23-149/mo	yes	yes	yes	17	US
Cryptohopper	Cloud platform	$29-129/mo	yes	yes	yes	16	US
PI Pionex	Exchange-native	0.05% per trade	yes	-	-	-	US
3Commas	Cloud platform	$15-140/mo	yes	yes	yes	9	US
HaasOnline	Cloud platform	$19.99-249.99/mo	yes	yes	yes	24	US
OKX Trading Bots	Exchange-native	free	yes	-	-	-	-

Live values from our database - details and caveats in each review. "-" = not documented.

Arbitrage is the strategy whose marketing survives entirely on the first line of that chart. The honest version: a gap between venues is an invitation to a race you enter with retail infrastructure against desks with colocated servers, negotiated fee tiers and pre-positioned inventory[7]. The gap closes while you read about it.

That does not make the features useless - it makes them tools for a specific setup: meaningful capital split across venues, fee tiers that shave the per-trade cost (BNB discount moves Binance to 0.075%[4]), and the ops discipline to manage inventory across exchanges. The platforms on this page sell the automation; the setup is what determines whether it pays.

Cross-exchange vs triangular vs market making

Cross-exchange (Bitsgap's spread model[1]): the intuitive one, and the transfer-bound one. Works best between venues with structurally different liquidity - and it must clear two fee schedules plus withdrawal costs per cycle.

Triangular (Cryptohopper's higher plans[2]): within one venue, no transfers, gaps measured in basis points and milliseconds[7]. On liquid BTC/ETH/USDT markets this is HFT territory; on thin pairs it is where retail sees "opportunities" that are really liquidity traps.

Market making (Hummingbot[6]): the professional's version of the same idea - quoted spreads on both sides, inventory-managed, no external transfers. Self-hosted, free, and as demanding as the self-hosted ops reality implies. If your interest is genuinely spread-shaped income, this is the lane that scales it honestly.

Tips that keep arbitrage from becoming a donation

• Compute the full stack per trade before subscribing. Taker fees both venues[4], withdrawal fee, estimated transfer minutes, slippage on both legs. If the screened spread does not clear the stack by a multiple, the screener is showing noise.
• Prefer inventory models on venues you already use. Capital split across two venues you trust beats transfer-dependent cycles - and the rebalance rule matters more than the trigger rule.
• Backtest the gap, not just the strategy. Bitsgap and 3Commas offer backtesting[1][3]; Hummingbot's paper mode[6] replays spread capture. A backtest without the fee stack is the first chart of this page wearing a suit.
• Watch withdrawal limits and maintenance windows - the transfer leg fails exactly when spreads widen. Calendar the venues' maintenance schedules; they are part of your edge or part of your losses.
• Treat any arbitrage pitch promising no downside as the scam it is. The FTC's marker[8], plus our scam patterns guide - arbitrage's vocabulary is the most abused in the space because the truth (thin, competitive, capital-heavy) sells badly.

Warnings

• Leg risk is the risk. When one leg fills and the other does not, you are directionally exposed exactly as the spread moves against you. Real arbitrage ops are leg-management ops.
• Fee tiers are the moat. Retail pays 0.1%[4]; desks pay less. Every comparison of "the same gap" across fee tiers ends the same way - the moat is why the desk eats first.
• Stablecoin "arb" pitches spike in crises. Depeg moments carry counterparty risk at its maximum - the arbitrage that pays most is priced in exactly the risk that kills it.
• Triangular gaps on illiquid pairs are traps. Thin books mean the second and third legs slip away from quoted prices - the screen shows the first leg's price only[7].
• Subscription + fees must clear captured spreads. $23-105/mo[1][2][3] against thin, competitive spreads is a losing equation below meaningful size - price it with the bot cost calculator before the subscription renews itself.

The inventory model, in operational detail

The version professionals run - and the version the cloud features approximate - pre-positions capital on both venues and never moves coins mid-trade. The operational reality behind the clean idea: capital splits (half idle at any moment by design), inventory drifts (the buying venue drains while the selling one fills), and rebalancing is itself a fee event that must be scheduled, sized and priced. Withdrawal fees and limits apply at rebalance; maintenance windows and withdrawal suspensions arrive exactly when spreads widen. None of this appears in feature lists because none of it is a feature - it is the job.

Add the venue risk dimension: pre-positioned inventory is unsecured exposure to both exchanges, and our /incidents/ timeline documents why that matters (Bybit 2025, Bitget 2026). Arbitrage capital is venue capital - the yield being chased competes with simply holding on a single regulated venue, which is a fair comparison the pitch never volunteers.

Stablecoin "arb", the special case of nothing

Peg deviations deserve their own paragraph because they generate the most seductive screenshots. A stablecoin trading at $0.97 looks like 3% for nothing; the honest reading is a 3% counterparty-risk quote - the market is pricing depeg or distress, and the "arbitrage" requires holding the instrument through exactly the event the discount prices. Routine peg noise (10-30 basis points) is smaller than the fee stack on most retail tiers[4]. The middle ground - small, persistent deviations on liquid pairs - belongs to the desks again.

The same logic scales to "arb" pitches that quote historical spreads: past gaps on dead pairs, screenshots without fee math, "proven" strategies with no methodology. The arbitrage vocabulary's history in this market is as a costume for fixed-return schemes[8] - which is why this page treats the honest version (capital-heavy, thin, operationally demanding) as the only one worth documenting.

If there is a defensible retail version, it is the boring one: two venues you already use, a pair with structural liquidity differences, inventory pre-positioned at sizes the position size calculator justifies, and a written rebalance rule. Run it on paper for a month, log every hypothetical fee and transfer, and let the ledger - not the screener - tell you whether the spread survives your stack. Most ledgers conclude the stack wins; that conclusion is worth $0 in subscriptions and a month of data. The rare exception - a structurally separated pair on venues you already fund - is exactly the kind of finding worth scaling slowly, one rebalance rule at a time. Document every cycle: entry spread, exit spread, fees both venues, transfer minutes. A month of that ledger is the only arbitrage curriculum that survives contact with real markets - and the cheapest tuition available.

What the numbers say

• 8 platforms in our tracking list arbitrage features; 75% offer backtesting - insist on gap-plus-fees backtests, not strategy-only.
• Median entry price across the subscription lanes: ~$23/mo/mo - the fixed cost under the variable stack.
• 2014 marks the oldest operator still listed in this lane - arbitrage tooling is old enough that its marketing cycles have repeated several times.
• The how bots work guide covers the execution loop arbitrage shares with every other bot - and why latency and rate limits[5] decide what retail can reach.