You can ask ChatGPT to write you an EA. It'll do it. Describe your strategy in plain English, get MQL5 code back in a minute.
But ask ChatGPT to predict whether the next candle on EURUSD H1 will be bullish or bearish, in real time, on every tick? It can't. It can't. That's not what it was built for.
There are two very different kinds of AI showing up in trading right now, and most people are confusing them. We use both at FXTool, for completely different jobs. This article is about telling them apart.
"AI" is not one thing
When people say "AI" in 2026, they usually mean ChatGPT. Or Claude. Or DeepSeek, Gemini, 豆包. Pick your favorite chatbot. These are all Large Language Models. LLMs. They read text, understand meaning, and generate text back. That's their core skill.
Then there's a different kind of AI that lives inside EAs. The kind you train in Python on historical price data, export as an .onnx file, and run locally inside MetaTrader 5. These are specialized ML models — small, fast, narrow. They don't understand language. They take numbers in, push numbers out.
Here's an analogy that actually works: an LLM is like a general practitioner. You can walk in with any question — symptoms, test results, concerns about your diet — and get a thoughtful answer. An ONNX model is like a blood pressure monitor. It measures one thing, instantly, with no opinion.
You wouldn't ask your blood pressure monitor why your knee hurts. And you wouldn't wait 3 seconds for your doctor to answer before deciding whether to take a trade.
What LLMs actually do for traders
If you've used ChatGPT or Claude for anything trading-related, you've probably done some of these:
Writing EA code. You describe a strategy — "buy when RSI crosses below 30 and the 50 EMA is above the 200 EMA, with a 50-pip stop loss" — and the model gives you compilable MQL5. Over 50% of trading systems now involve some AI assistance in development.
Analyzing news and sentiment. Paste in a central bank statement, an earnings report, or a batch of tweets. The model reads it and tells you the tone — hawkish, dovish, neutral — with a confidence score. Some traders feed this into their decision process.
Explaining backtests. You ran a backtest, got a report full of numbers, and you're not sure what they mean. "Is a 12% max drawdown on a gold scalper normal?" An LLM can contextualize that for you. We've written about what backtest reports actually tell you if you want the non-AI version.
Learning and debugging. "Why is my EA opening double positions?" Paste the code, get an explanation. It's the fastest way to learn MQL5 right now, honestly.
The limitation is fundamental though. LLMs live in the cloud. Every question you ask goes over the internet, gets processed on someone else's server, and comes back. That takes 1 to 5 seconds per request. And you pay per token.
For writing code or analyzing a document, that's fine. For making a trade decision on every candle or every tick? Way too slow, way too expensive.
What ONNX models do inside EAs
ONNX models solve the opposite problem. They don't understand language at all. They can't write code or explain anything. What they do is: take a fixed set of numbers (price data, indicator values, volatility measures) and output a prediction. Buy, sell, or hold. A probability. A price target.
And they do it in microseconds, running locally on your machine, inside the EA process itself. No internet. No API call. No cost per query.
A typical setup: you train a model in Python that takes the last 120 H1 candles of OHLC data and predicts whether the next candle closes higher or lower. Export it as .onnx. Load it in your EA with OnnxCreate(). On every new bar, feed in the latest data, call OnnxRun(), get back a number. That number drives your trade logic.
We covered the full workflow in our previous article on ONNX in EA development.
The limitations are also real. ONNX models:
- Can't read news. They only see numbers you give them.
- Can't explain why. You get a prediction, not a reason.
- Need training data. Someone has to build and train the model first.
- Go stale. Market structure changes, the model degrades, you retrain.
If you've read about overfitting and curve fitting, you know the training part is where most people go wrong.
The comparison, in one table
| LLM (ChatGPT, Claude) | ONNX model | |
|---|---|---|
| What it does | Understands language, generates text | Numbers in, numbers out |
| Speed | 1-5 seconds per request | Microseconds |
| Where it runs | Cloud (API call) | Locally inside the EA |
| Cost | Pay per token | Free after training |
| Can explain its decisions | Yes | No |
| Can read news | Yes | No |
| Can make real-time trade decisions | Too slow | Yes, that's the point |
| Typical parameter count | 100 billion+ | A few million |
| Needs internet | Yes | No |
Source for the speed comparison: MQL5 community benchmark of ONNX vs LLM approaches in MT5.
The smart play is using both
Some developers on MQL5 are already running hybrid systems. An LLM handles the macro layer: processing news, gauging sentiment, deciding whether today is a good day to trade at all. The ONNX model handles execution. On each bar, should we enter, exit, or wait?
One MQL5 developer built exactly this: DeepSeek for reasoning and sentiment, a compiled ML model for execution timing. Two AIs doing two different jobs in the same EA.
We do something similar at FXTool, though not in a single EA. Claude helps us write and debug EA code, analyze strategy ideas, and produce content. ONNX models run inside the EAs we test, doing the actual prediction work. Different tools, different jobs, same team.
What this means for you
You don't need to understand the technical details. Just keep this straight:
If you want AI to help you write, fix, or understand an EA, use ChatGPT or Claude. Describe what you want in plain language. There's a guide to AI-assisted EA coding if you want to get started.
If you want AI running inside your EA making real-time predictions, that's an ONNX model. Different technology, different workflow. You need Python, training data, and some ML knowledge. Or find an EA that already has one built in.
Not sure where to begin? Start with an LLM. Ask it to explain what an EA is, how to choose one, or how to read a backtest report.
LLMs and ONNX models live in the same toolbox, but they're not the same tool. One is a Swiss army knife: versatile, handles anything, master of none. The other is a scalpel. Does one thing, does it fast.
You wouldn't use a scalpel to open a can. And you wouldn't wait for a Swiss army knife to take your blood pressure.
About the author: The FXTool team builds and tests MetaTrader trading tools daily. We run every EA we sell on live accounts and publish the results. This guide reflects what we've learned from building 50+ EAs and working with thousands of retail traders.