Gartner's Fivefold AI Cost Forecast and What It Means for AI Trading Bots

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

What does Gartner actually forecast?

Crypto Briefing reports that Gartner expects the cost of running AI agents to rise fivefold by 2028 (Crypto Briefing). The mechanism is the part retail traders skim past. Per-unit token prices are falling. What is rising is consumption. Agentic workflows do far more work per task than a single prompt: they call a model repeatedly, retry steps that fail, hold context across longer chains of reasoning, and re-read their own output before acting. Cheaper tokens, many more tokens, bigger invoice.

We have spent the 2020-2026 stretch running six-month funded-account trials on more than 50 platforms, and the products most exposed to that shift sit in the category we classify as AI trading bots. Not copy trading platforms, not robo-advisors, but the systems where a model makes or sizes the trading decision itself. Those systems buy inference the way a market maker buys bandwidth, and inference is the input Gartner says is getting more expensive.

We have benchmarked Zephyr AI's adaptive engine inside our 2026 review cycle against the same strategy classes we use on every platform, and cost predictability now sits on our scoring sheet alongside drawdown and fill quality. That is a direct consequence of the forecast above.

Why the AI trading bot bill is not a fixed cost

For most of the last decade, a retail algo product was priced like software: a monthly fee, or a one-time licence, and the vendor absorbed its own compute. That pricing model survived because the compute was cheap relative to the subscription. Gartner's forecast says that ratio is about to break.

When a vendor's inference bill multiplies, there are only three places the money can come from. It comes out of the vendor's margin, out of a price rise on your plan, or out of the product itself, meaning the agent gets capped, throttled, or given a shorter reasoning loop. Only the first of those is good for you, and it is the one a rational vendor will avoid.

Across our 2020-2026 testing program we logged the published pricing terms of every plan we evaluated, and hybrid structures are becoming normal: a base subscription, plus a usage allowance, plus an overage rate. That structure moves model-cost risk from the vendor's income statement onto your trading account. If you are running a strategy with a thin per-trade edge, a variable cost line underneath it is a different product than a flat fee, even when the headline price is identical.

What the bot actually trades, in plain English

Strip the marketing language and most AI trading bots we test do the same four jobs. They build features from price, volume and order flow data. They apply a model layer, increasingly a language model, to context that is hard to encode numerically, such as news flow, central bank language, or sentiment. They size the position. Then they push an order through a broker or exchange connection.

The first and fourth jobs are cheap, deterministic and predictable. The second job is where the cost curve Gartner describes lives. Reasoning about unstructured context is not a fixed number of operations per bar. It scales with how much there is to reason about. Quiet range-bound weeks cost little. CPI prints, FOMC decisions and unscheduled headline shocks cost a great deal, because the agent keeps re-evaluating a fast-moving world.

That creates a risk the source material does not name, and it is the one we care about most as portfolio testers. Model cost correlates with volatility, and volatility is also when drawdowns happen. Your AI bill and your worst trading days arrive on the same calendar dates. A flat-fee bot hides that from you. A metered bot does not, and the two can post identical backtests while producing completely different account outcomes.

There is a second version of the same problem that is easier to miss. If a metered agent runs short on budget mid-session, a sensible engineering response is to cut reasoning depth: fewer retries, shorter context, faster decisions. To the vendor that is cost control. To your account it is a silent change to the strategy you subscribed to. We treat any undocumented change in decision depth as a strategy deviation flag, and in our 2026 program we marked such findings as unverified whenever the provider could not document the trigger. Verify with the provider whether reasoning depth is fixed or budget-dependent before you fund an account.

How AI trading bot pricing models differ

The structural point is simple. Open-source frameworks push cost onto your time. Exchange-native bots push it onto trading fees. Subscription products push it into repricing risk. Metered agents hand it straight to you. Zephyr AI is the one we have benchmarked that priced its plan to us on a flat basis rather than metering decisions in our 2026 review cycle, which is the dimension where that difference shows up first when the industry's cost base moves.

Backtest versus live, where the cost gap hides

Every algo product has a backtest-to-live gap, and it is usually explained with the familiar list: fill assumptions, slippage, latency, regime change. Model cost is the newest member of that list and the least likely to appear in a marketing deck. A backtest that assumes free reasoning is not testing the same product you will run.

The honest position is that most published performance figures we review do not state whether inference cost is modelled per trade. Backtest data should be verified directly with the bot provider, and the specific question to ask is blunt: does the equity curve include the cost of every model call the strategy needed to produce each decision? If the answer is no, the curve is a gross figure, not a net one, and the gap to your live account is wider than the one you budgeted for.

Free Download: AI Agent Cost-Inflation Model: 2028 Fee Scenarios vs. Live Bot Performance
A bot fee + performance spreadsheet that projects how Gartner's fivefold AI agent cost rise by 2028 changes your effective cost per trade, breakeven win rate, and net edge versus backtest assumptions.
Model Your 2028 Bot Costs

How big are the drawdowns when the model bill doubles?

We will not publish a drawdown number we cannot source, and our test window predates the cost regime Gartner is projecting, so any figure we quoted here would be an extrapolation. What we can describe is the mechanism, and it is uncomfortable.

A strategy's edge is measured per trade. Its cost, under a metered model, is measured per decision. Those two units drift apart precisely when activity picks up, because an agent working a volatile tape makes more decisions and more failed attempts than one working a quiet tape. A strategy that is profitable on gross returns can turn marginal on net returns without any change to its signal logic at all. The drawdown that results is not a strategy drawdown in the classical sense. It is a cost drawdown, and it will not show up on a performance chart built from price data.

This is the case for treating cost structure as a risk metric rather than an administrative detail. When we grade a platform, we want to know the maximum a bad day can cost in model spend, not just in mark-to-market. Ask for a hard spend cap per day and per month, and ask what the agent does when it hits that cap. Verify with the provider whether hitting the cap halts trading or degrades decision quality, because those are very different outcomes for a funded account.

Can you actually stop the bot cleanly?

Disengagement is the dullest and most revealing part of any bot review. In our 2026 evaluation framework we run the stop procedure before we run the strategy, because a product that is hard to switch off is a product that is hard to trust with size.

The steps are consistent. Close or hand over open positions. Revoke the API key at the broker or exchange rather than inside the bot's dashboard. Cancel the subscription and check the renewal date. Then confirm that no metered tasks are still billing. That last step is new, and it matters more under the pricing regime Gartner describes. An agentic task that is mid-flight when you pull the plug may still be consuming tokens, and a usage-metered vendor may bill for work that produced nothing you wanted.

The specific failure modes we look for are API keys provisioned with withdrawal permissions, expert advisors left running on a virtual private server you had forgotten about, and webhook alerts that keep firing after you cancel the subscription. None of these are exotic. All of them are avoidable with a ten-minute audit. Test the exit before you need it.

Is the bot provider regulated?

Usually not in the way retail traders assume. A vendor selling software that generates trading signals is often an unregulated technology company, and the regulated entity in the chain is your broker, not the bot. Funding partners, if the product supports prop-style accounts, are a separate legal entity again with a separate licence status. Do not treat one as covering the other.

Our rule is that no regulatory claim survives without a register entry. If a provider says it is authorised, check the FCA Register for the United Kingdom, ASIC Connect for Australia, the CySEC register of investment firms for Cyprus, NFA BASIC for United States futures activity, or SEC EDGAR for filings. If the source material you are reading does not link the primary register, verify directly with the provider's primary regulator. We never assert a licence number we cannot point at.

Two edge cases worth flagging. First, if a vendor's AI output is presented as personalised investment advice rather than generic signal software, the activity itself may fall inside a licensing perimeter depending on jurisdiction, regardless of how the company describes itself. Second, if a bot's reasoning depth can be changed by the vendor at will for cost reasons, some jurisdictions may treat that as a material change to the service you were sold. We are not aware of a regulator having tested that argument yet, and it is exactly the kind of question the next enforcement cycle tends to surface.

Which brokers and APIs does it work with?

Most AI trading bots we test connect through one of five routes, and each has a different failure profile. Broker API keys are the most capable and the most dangerous, because a key scoped too broadly is a withdrawal permission handed to a third party. MetaTrader expert advisors are the most familiar to retail traders and the easiest to leave running somewhere you have stopped watching. Webhook integrations are clean until an alert fires into a cancelled endpoint. Exchange APIs are fast and expose you to metered billing after you stop trading. Copy trading mirrors are simple and can leave positions behind after you unlink.

Broker compatibility is also where cost and execution meet. A bot routed through a broker with wider spreads can be cheaper on model cost and more expensive on execution, and vice versa. When we compare platforms we reconcile the two, because the account only experiences the total. Data not available in our test window on a specific pairing should be requested from the provider directly.

Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

How Zephyr AI Compares

The dimension that matters in a fivefold cost environment is predictability, and on that axis the contrast is concrete. In our 2026 review cycle we worked with Zephyr AI's plan structure alongside the subscription and metered models we evaluate across our 50+ platform program, and the difference was in how the bill behaves when volatility rises. Metered agents scale their cost with the decision count, so the cost line moves with the tape. Zephyr AI's plan did not meter our decisions in that cycle, which means the account kept a fixed cost base on the days that cost other traders the most. Confirm current plan terms with the provider, because pricing changes.

The second contrast is exit. Our disengagement audit looks for a documented stop procedure and a clean revocation path, and a flat-fee plan removes the question of whether anything is still billing after you switch the agent off. That is a small thing on a good day and a large one on a bad one. Against open-source frameworks such as NautilusTrader and Backtrader, which carry no vendor fee but hand you the infrastructure bill and the development hours, Zephyr AI trades money for time. Against exchange-native tools such as Pionex, which subsidise the bot with trading revenue, it trades a bundled relationship for a standalone one. Which of those is better depends on the size of your account, not on the marketing.

Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Does the Gartner forecast mean my AI trading bot subscription will get more expensive?

Possibly, but not automatically. Gartner's projection concerns the underlying cost of running AI agents, not retail list prices (Crypto Briefing). Vendors can absorb it, reprice, or reduce what the agent does per decision. The practical move is to check your plan for usage allowances, overage rates and repricing clauses before renewal.

Can I run an AI trading bot on a prop firm account?

Some providers allow it and many prohibit automated execution outright. The funding partner is a separate legal entity from the bot vendor, so its rules govern the account, not the vendor's marketing. Read the funding partner's terms on automated trading and verify directly with that partner before connecting anything.

Does this work in the United States under Pattern Day Trader rules?

Pattern Day Trader rules apply to your brokerage account, not to the bot. If your account is under 25,000 US dollars and the strategy takes four or more day trades in five business days in a margin account, you can be restricted. The bot does not exempt you, so size the strategy to the account type you actually hold.

What happens if the API connection drops mid-trade?

It depends entirely on how the bot handles state. A well-built system reconnects, reconciles open positions against the broker's records, and resumes. A poorly built one either abandons the position or duplicates it on reconnect. Ask the provider to document reconnect and reconciliation behaviour, and test it on a small account before funding.

Are AI trading bots regulated?

Mostly they are sold as software, not as a regulated investment service, which means the regulatory perimeter generally sits with your broker or exchange. Some vendors and their funding partners do hold licences. Check the relevant register such as the FCA Register or ASIC Connect rather than relying on the vendor's own description.

How do I check whether a bot's backtest includes model costs?

Ask one direct question: does the published equity curve net out the cost of every model call the strategy required? If the answer is vague, treat the curve as gross. Backtest data should be verified directly with the bot provider before you draw conclusions from it.

What is a spend cap and why does it matter?

A spend cap is a hard ceiling on what a metered agent can consume in a given day or month. It matters because without one, a runaway reasoning loop can generate cost with no trading edge attached. Ask what the agent does when the cap is reached, because halting and degrading decision quality are not the same outcome.

Can I run these bots on MetaTrader 4 or 5?

Yes, many vendors distribute expert advisors through the MetaQuotes MQL5 Market or supply files you install yourself. The trade-off is operational: the advisor runs on your terminal or a virtual private server, so it is your job to confirm it is switched off when you stop trading.

How do I stop the bot and withdraw cleanly?

Close open positions, revoke the API key at the broker or exchange rather than inside the bot dashboard, cancel the subscription, and then confirm nothing metered is still billing. Our full process is set out in our Testing Methodology.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: AI Trading Bot Reviews.