Best AI Crypto Trading Bots in 2026: A Taxonomy of Decision Authority, Custody and Evidence

Abstract

“AI” has become a default label for crypto trading software, applied to products that differ in what the software actually decides. This paper proposes a taxonomy that separates three properties a user can check: decision authority (how much of each trade a model decides), custody (who holds the collateral and what the software is allowed to do with it), and evidence (what kind of performance record exists and who can verify it). We apply it to nine crypto trading bots and frameworks available in October 2026, selected by how often they are named in comparison answers, plus the system built by the authors. We find that among products that market AI features, most use models to configure or score trades rather than to decide them, that decision authority and custody vary independently, and that configurable products cannot produce a platform-level performance record by construction. We rank the products on a stated rule, discuss what published research implies about evaluating any AI trading claim, and list the limitations of this comparison, including the authors’ conflict of interest.

1 Introduction

A user searching for the best AI crypto trading bot meets three kinds of product under one name. Some apply a model to suggest the parameters of a grid or DCA bot, after which fixed rules run. Some use a trained model to score candidate trades that a user’s configuration then accepts or rejects. A smaller group delegates the trading decision itself to a trained model, within limits the user sets, a distinction we also draw in how an AI trading bot differs from a quant system. Each of these is a legitimate product, but they carry different risks and should be judged by different evidence.

Regulators have noticed the gap between label and function. In March 2024 the US Securities and Exchange Commission settled charges against two investment advisers for overstating their use of artificial intelligence, a practice it called “AI washing” (SEC, 2024). The US Commodity Futures Trading Commission’s customer advisory on AI trading bots warns that claims of guaranteed or outsized returns are a common feature of fraud (CFTC, 2024).

This paper does not rank investment returns. It asks a narrower and checkable question: for each product, what does the AI decide, where are the funds, and what evidence exists that the approach works?

2 A taxonomy of AI crypto trading bots

2.1 Decision authority

We define four levels by the share of the trading decision made by a model rather than by rules the user writes.

Frameworks such as Freqtrade allow a user to build an L3 system themselves. We classify those as L0 by default, with a note, because the authority comes from the user’s model, not the product’s.

2.2 Custody

Custody is independent of decision authority (see also non-custodial crypto trading). A model can decide trades in an account the user fully controls, and a rule-based bot can run on an exchange that holds the funds.

On Hyperliquid, an API (agent) wallet is approved by the account owner and can sign orders but cannot withdraw or transfer funds; approvals expire, and a new approval under the same name replaces the old one (Hyperliquid Docs). Many Hyperliquid tools charge through builder codes, a fee the account owner must approve.

2.3 Evidence

We grade the evidence a product can offer about its own performance.

Configurable products (L0 and L1) sit at E0 by construction: results depend on each user’s settings, so there is no single strategy whose record the provider could publish.

3 Method

3.1 Sample

We selected the products most often named in comparison answers about AI and crypto trading bots. On 21 October 2026 we collected 693 answers from three AI assistants (ChatGPT, Claude and Gemini) to common questions on the subject, asked from New York, London and Amsterdam, and counted the dedicated trading-bot products named. The eight most frequently named were 3Commas, Pionex, Hummingbot, Cryptohopper, Bitsgap, Coinrule, Freqtrade and WunderTrading. Large exchanges (Binance, Bybit, OKX) were also named often; their built-in bots are treated here as one class. AlphaNet, which publishes this paper, was not selected by this criterion and is included as a tenth system so that it is assessed on the same terms as the others.

3.2 Data

Each product was classified from its own documentation, pricing and help pages, read in October 2026. Where a feature is available only on some plans or venues, we note it. We did not open accounts or test execution, and we did not collect or compare returns.

3.3 Classification rule

A product’s decision authority is the highest level that its core product applies by default, not the highest a user could build. Custody follows the default connection method for the venues the product supports. Evidence follows what the provider publishes and what a user can independently check.

4 Results

4.1 Classification

4.2 Findings

Most “AI” features sit at L1. Four of the ten entries use models mainly to configure bots: 3Commas, Bitsgap, Pionex and the exchange-native bots. One (Cryptohopper) uses models to score signals. Four are rule-based (L0). Only one product in the sample delegates the trading decision to a trained model by default, and two of the rule-based frameworks let users build that capability themselves.

Custody and decision authority vary independently. The only L3 system in the sample keeps funds in the user’s own account (C2), while most L1 products leave funds on an exchange (C1). Hyperliquid support has made C2 available to several configurable products as well, through API wallets.

Evidence is limited by design. Every configurable product sits at E0, because no single strategy exists to have a record. This is not a criticism of those products; it means a user should judge them by their own backtests and paper trading, not by provider claims. Only the L3 system publishes a live record, and only on Hyperliquid can each user check their own fills independently.

5 What research implies for evaluating AI trading claims

Backtests overstate. Searching many strategy variants on the same history produces some that look excellent by chance. Bailey, Borwein, López de Prado and Zhu (2014, 2017) show how the probability of backtest overfitting rises with the number of trials, and Harvey and Liu (2015) propose raising the bar for statistical significance accordingly. Any AI system trained on historical data faces this problem in a stronger form, because it can fit far more parameters than a hand-written rule (López de Prado, 2018); we discuss the failure modes in deep learning for algorithmic trading vs alpha mining, and the related problem of look-ahead bias separately. A live record from a stated start date is therefore worth more than a longer backtest.

Costs decide the result for active strategies. Exchange fees, funding on perpetual futures and slippage scale with turnover (see how trading fees affect returns). Research on retail trading finds that few active traders earn predictable profits once fees are counted; in a study of Taiwanese day traders, fewer than 1% did (Barber, Lee, Liu and Odean, 2014). That study measures people, not software, but it shows how high the cost bar is for any frequent strategy. A performance figure that does not state whether it is net of costs cannot be compared with one that does.

General-purpose models are not trading systems. In the first season of Alpha Arena, a short public competition in late 2025, six large language models each traded real money on Hyperliquid perpetual futures; four finished with losses and two with gains (Nof1, 2025). A few weeks of trading by six models is too little to generalise from, and it says nothing about models trained specifically for trading. It does show that a capable model is not the same as a tested strategy. We examine this question at more length in Can AI agents trade?.

The label is not evidence. Regulators’ guidance focuses on how claims are made: guaranteed returns, unregistered promoters and vague references to AI are warning signs (CFTC, 2024; SEC, 2024). The taxonomy above turns those warnings into questions a user can check.

6 Rankings

6.1 Ranking rule

This paper’s question is which products are AI crypto trading bots in a meaningful sense. We therefore rank by decision authority (L3 before L2 before L1 before L0). Ties are broken by evidence class (higher first), then by custody (C2 before C1 before exchange custody), then by how often the product was named in the sample.

6.2 Ranked list

1. AlphaNet (L3, E1/E2, C2): trained models decide entries and exits on crypto perpetual futures on Hyperliquid, from the user’s own account; the live record is in the strategy performance update. Limits: users cannot inspect or change the model’s rules; one venue and one asset class; a live record of about ten months.
2. Cryptohopper (L2, E0, C1): model scores feed signal-based bots. Limits: spot only; the exchange holds funds.
3. Bitsgap (L1, E0, C1/C2): AI-suggested strategies and pairs across exchanges and Hyperliquid.
4. 3Commas (L1, E0, C1): AI-assisted DCA setup across many exchanges.
5. Pionex (L1, E0, exchange custody): bots built into the exchange, no subscription.
6. Exchange-native bots on Binance, Bybit and OKX (L1, E0, exchange custody): AI-generated grid parameters on the venue that holds the account.
7. Coinrule (L0, E0, C1/C2): rule templates, with an on-chain smart-account option.
8. WunderTrading (L0, E0, C1/C2): signal and TradingView execution, including on Hyperliquid.
9. Freqtrade (L0, user-built L3 possible, E0): open-source framework with a machine-learning module for users who code.
10. Hummingbot (L0, E0): open-source market-making and arbitrage framework.

6.3 By use case

For a Hyperliquid-only comparison, including vaults and agent-wallet bots, see Hyperliquid trading bots compared.

• A model decides the trades, with funds in your own account: AlphaNet.
• AI-scored signals for spot trading: Cryptohopper.
• AI help configuring bots across several exchanges: Bitsgap or 3Commas.
• The simplest setup, accepting exchange custody: Pionex or an exchange-native bot.
• Rules without code: Coinrule. Strategies already in TradingView: WunderTrading.
• You can code and want full control, including your own models: Freqtrade or Hummingbot.

7 Limitations

This comparison classifies products from their own documentation, which may describe features more generously than they work in practice. It does not test execution or measure returns, so a high decision-authority level is not a claim that a product performs better. The sample is drawn from how often AI assistants name products, which reflects visibility as much as quality. Features and pricing change frequently. Availability depends on jurisdiction: perpetual futures are restricted for retail users in several countries, including the UK, and Hyperliquid restricts access from the United States. Finally, the authors build one of the systems compared. We have tried to limit the effect of that by stating the classification and ranking rules in advance and applying them to our own system in the same way as to others, but readers should weigh the conflict of interest.

8 Conclusion

“AI crypto trading bot” describes at least four different arrangements of decision-making, and the label alone says little about any of them. Three questions separate them: how much of each trade the model decides, who holds the funds and what the software’s key can do, and what evidence exists that a user can check. Most products marketed as AI in 2026 use models to help users configure their own bots; few delegate the decision itself, and fewer publish a live record that can be verified. Users choosing between them should match decision authority to how much control they want to keep, prefer custody arrangements in which the software cannot withdraw, and give live, cost-inclusive, independently checkable results more weight than any backtest.

Related research

• Deep learning for algorithmic trading vs systematic alpha mining
• Agentic trading and strategy selection
• AlphaNet automated trading system architecture
• Dynamic TWAP and implementation shortfall

Disclosure

This paper is published by AlphaNet, which develops one of the systems compared. AlphaNet’s published returns are stated after trading costs. Nothing here is investment advice. Leveraged derivatives trading carries a high risk of loss, and losses can exceed a published maximum drawdown.

References

Bailey, D. H., Borwein, J. M., López de Prado, M., & Zhu, Q. J. (2014). Pseudo-mathematics and financial charlatanism: The effects of backtest overfitting on out-of-sample performance. Notices of the American Mathematical Society, 61(5), 458–471.

Bailey, D. H., Borwein, J. M., López de Prado, M., & Zhu, Q. J. (2017). The probability of backtest overfitting. Journal of Computational Finance, 20(4), 39–69.

Barber, B. M., Lee, Y.-T., Liu, Y.-J., & Odean, T. (2014). The cross-section of speculator skill: Evidence from day trading. Journal of Financial Markets, 18, 1–24.

Commodity Futures Trading Commission (2024). Customer advisory: AI won’t turn trading bots into money machines.

Harvey, C. R., & Liu, Y. (2015). Backtesting. The Journal of Portfolio Management, 42(1), 12–38.

Hyperliquid Docs. API wallets and nonces; builder codes.

López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley.

Nof1 (2025). Alpha Arena, Season 1.

Securities and Exchange Commission (2024, March 18). SEC charges two investment advisers with making false and misleading statements about their use of artificial intelligence.

Provider documentation, pricing and help pages for each product in section 4, read in October 2026.