Every few weeks a new model becomes “the best ever”, and this week it’s Google’s turn with Gemini 4 Argon. If you run a Gemini trading bot, or any AI EA, you’re probably asking the practical question: do I switch now, wait, or is my whole setup already outdated?
I’ve been through a lot of these launches with a live account connected to the models. The headline benchmark has never been the thing that decided what I left running. Here’s what Google actually announced, what my live record says about Gemini, and how to test Argon on your own chart the day it opens.
What Google announced about Gemini 4 Argon (and what it didn’t)
Google announced Gemini 4 Argon on 30 September 2026. At the time of writing (1 October), this is what’s public:
- Limited rollout first. Argon is going to a group of trusted cyber defenders through Google’s Fairwind Program. Broad availability is planned “as soon as possible”, starting with paid API customers and Google AI Ultra subscribers.
- Introductory pricing: $2 per million input tokens and $10 per million output tokens, with a higher standard rate announced for later.
- Bigger output: the output limit goes up to 1M tokens.
- Benchmarks: Google says Argon scores above GPT-6 Astra, Claude Opus and Fable on several benchmarks. The same reports note GPT-6 Astra still leads some software and computer-use tasks, and Claude Opus 5.5 leads in terminal-agent work.
What it didn’t announce: an API model ID you can use today, or a single word about trading. Every number above was produced on tasks that have nothing to do with a gold chart at 14:30 on a Friday.
Benchmarks don’t place trades
An EA doesn’t need the smartest answer once. It needs an acceptable answer thousands of times, on time, in the format the code expects. When I judge a model for live trading, these are the things that actually matter:
- Stability. Does it answer every request, or does it time out and fail when the market is busy?
- Latency. A decision that arrives after the setup is gone is worth nothing.
- Clean output. The EA has to parse the response. A brilliant answer in the wrong format is a skipped trade.
- Consistency. The same context should get a similar decision. A model that flips its view on identical data can’t be audited.
- Cost per day. Your API bill keeps running whether the trade wins or not.
None of those show up in a launch chart. You find them on a demo account over a few weeks, which is exactly why I don’t switch models on launch day.
The quiet winner in my live record: Gemini 3.1 Pro
This spring Gemini 3.1 Pro was the model on my live Alpha Pulse account, and for a couple of months I ran it side by side with Claude Opus 4.7 on the same XAUUSD account. Same prompts, same risk, same broker. I wrote up the full comparison in Opus 4.7 vs Gemini 3.1: live forex EA comparison.
The honest summary: Opus reasoned better when a setup was ambiguous. Gemini 3.1 Pro was faster and steadier when the market was busy, and it kept delivering decisions when other models were struggling. For an EA that runs all week without you watching it, steady beats brilliant.
So if you ask me which model I’d leave running unattended for a month, my answer is still a Gemini Pro. Not the models that got all the attention. The hype went to Opus and Fable; Gemini just kept doing the job.
That’s also why Argon is interesting to me. If Google kept the stability of the 3.1 generation and added the reasoning the benchmarks claim, that’s the combination I’ve wanted for a long time. “If” is doing a lot of work in that sentence, and only a live test removes it.
Running a Gemini trading bot in MT5 the day Argon opens
You don’t need to wait for a new EA version to try a new model. In Alpha Pulse AI, this is the route:
- Wait for the real model ID. Google hasn’t published Argon’s API ID yet. When it appears in Google’s API documentation, copy it exactly.
- Select Custom Model. In the inputs, set AI Model Selection to Custom Model and paste the ID into the Custom Model field. The EA detects Google from “gemini” in the name; you can also write it with a
google:prefix. - Use your Google API key. The same key you use for the Gemini presets. Argon may need a paid tier at first, so check your access before you start.
- Check WebRequest.
generativelanguage.googleapis.commust be in MT5’s allowed URLs. If you already run a Gemini preset, it is. - Run it on demo, not on your live account. MT5’s Strategy Tester can’t do this: the API call needs WebRequest, which the tester doesn’t allow. Forward test only.
Gemini 3.8 Flash is the default preset today. Models that leave the preset list also stay usable this way, by typing their ID, so an older model you trust doesn’t disappear just because something newer arrived.
Watch the pattern behind every launch: in GPT-5.5 Is Here. Fake AI Trading Bots Are Coming Too, I explain why every new model brings a wave of “AI bots” that never connect to the model at all. Argon will bring the same wave.
My first-week test plan for Gemini 4 Argon
When Argon opens, this is what I’ll record on demo next to my current model. Copy it if you want to run the same test:
| What I record | Why it matters |
|---|---|
| Failed or timed-out requests per day | Stability is the first thing that breaks a live EA |
| Response time on news-heavy sessions | Latency decides whether the decision is still tradeable |
| Responses the EA couldn’t parse | Every one of those is a skipped decision |
| Agreement with my current model | When two models agree with high confidence, that signal has been worth more to me than either one alone |
| Decisions to wait, not only entries | A model that never waits is a model that overtrades |
| API cost per day | Introductory pricing won’t last forever |
Notice what’s missing: profit. One week of trades tells you nothing about an edge. It does tell you whether the model is stable enough to deserve a longer test. If you’ve never set up a forward test record, start with what to record in an AI trading forward test.
If your longer-term goal includes trading allocated capital, keep that decision separate from the model question. My Axi Select overview explains how that programme works. Permitted automation needs its own check there. This is an affiliate link; I may earn a commission if you sign up through it.
Gemini vs Claude for trading: should you switch?
Here’s how I’d decide, depending on where you are:
- Your current model has a record you trust. Keep it live. Test Argon on demo in parallel and compare your own numbers after a few weeks.
- You’re on Claude and happy with the reasoning. Same rule. Opus earned its place on ambiguous setups; don’t throw that away for a launch chart.
- You’ve had stability problems. This is where a Gemini Pro model is worth testing first, and Argon goes to the top of that list once it opens.
- You haven’t started yet. Don’t start with the newest model. Start with a stable one, build your record, then test new models against it.
The models will keep changing every few weeks. Your process is what you keep: the prompt, the risk rules, the record and the habit of testing before trusting. That’s the part I’d invest in.
Capital programmes deserve the same patience. Before planning around one, read the requirements and costs next to the opportunity. The Axi Select breakdown helps you make that decision on its own terms.
If you want to test models like this inside MetaTrader, have a look at DoIt Alpha Pulse AI for MT5. Read the features and setup requirements, check the current price and API costs, and decide if it fits the way you want to trade.
I’ll share my first Argon notes as soon as I’ve run it on demo. If you want them, join the DoItTrading newsletter.
Trading with AI EAs involves real risk to capital. Past results of any model or configuration don’t guarantee future ones.
Frequently Asked Questions
Is Gemini 4 Argon available in the API?
Not for everyone yet. On 1 October 2026 Google was rolling it out to a limited group first and said paid API customers and Google AI Ultra subscribers come next. Check Google’s API documentation for the current status and the exact model ID.
Can I use Gemini as a trading bot in MT5?
A model can’t trade on its own. It needs an EA that sends it market data and executes its decisions. Alpha Pulse AI does that with Gemini presets and a Custom Model input for any model ID.
Is Gemini better than Claude for trading?
In my live runs, Gemini 3.1 Pro was faster and steadier, while Claude Opus 4.7 reasoned better on ambiguous setups. Which matters more depends on your strategy, so test both on demo with the same prompt and risk.
Do I need a new EA version to use a new model?
Not in Alpha Pulse AI. Select Custom Model and type the model ID once the provider publishes it. The EA detects Google models from the name.
Can I backtest a Gemini trading bot in Strategy Tester?
No. The API connection relies on WebRequest, which MT5’s Strategy Tester doesn’t allow. Forward test on a demo account and keep a record of the configuration.
