Opus 4.7 vs Gemini 3.1: Live Forex EA Comparison (Real Trades)

Q: Can I really run both models on the same EA?

Yes, that's the design of DoIt Alpha Pulse AI . You bring API keys for Anthropic, Google, OpenAI, xAI, DeepSeek or Qwen and switch in the panel. There's no model lock-in. More setup questions answered here.

Most “AI trading model comparisons” you’ll find online were written by someone who tested each model for ten minutes on a backtest. I’ve been running Claude Opus 4.7 and Gemini 3.1 Pro on the same live XAUUSD account for the last couple of months. Same prompts, same risk, same broker. This is what actually showed up.

Why this comparison is different from the ones you’ve already read

Open YouTube right now and search “Opus 4.7 vs Gemini 3.1.” You’ll get reasoning benchmarks. Coding benchmarks. Math benchmarks. Tons of side-by-side prompts with cherry-picked answers.

None of that tells you what these models do when there is real money on the screen and the next decision could lose €200.

That is a completely different problem from “write me a poem about Tuesday.” It involves uncertainty, partial information, asymmetric risk, and a hard time limit. Two models that score identically on a public benchmark can behave very differently when you wire them into a forex EA and let them decide whether to take a XAUUSD setup at 14:32 on a Friday with US news in 18 minutes.

I run both of them, on the same live forward-test account, every week. The decisions are public on the Myfxbook. The video sessions are public on my YouTube channel. So instead of giving you another benchmark recap, I’m going to walk you through the patterns I keep seeing — the ones nobody talks about because nobody actually runs both with real fills.

If you’re considering an AI-driven EA and you want to know which model to use, this should save you a lot of API budget.

Setup: how I run Opus 4.7 and Gemini 3.1 Pro on the same EA

Before the comparison makes any sense, the methodology matters. Otherwise you get a “Claude is smarter” or “Gemini is faster” take that depends entirely on the prompt the writer happened to throw at it.

Here is how the test runs:

Same EA: DoIt Alpha Pulse AI — the LLM-decision-layer EA I built. It accepts any model with an API key. You swap providers in the panel; the rest of the system stays identical.
Same instrument: XAUUSD on M15.
Same broker, same account: live forward test on Axi Select, RoboForex setup variant for cross-validation.
Same risk per trade: ~0.5% on the live forward, lower than I’d run on a personal account.
Same prompt structure: the system prompt that ships with Alpha Pulse, with the same context window (price action levels, news pulse, recent volatility, account state).
What changes: only the model. Opus 4.7 one week, Gemini 3.1 Pro the next, occasionally side-by-side on alternating signals.

If you want to see the live decisions instead of reading my summary, the Myfxbook is open and the live XAUUSD sessions are on YouTube every week. The Alpha Pulse landing has both linked from the live results section.

↗

Want the EA that actually lets you swap models like this?
Bring your own API keys. Swap Opus, Gemini, GPT, Grok in seconds. Audit every decision.
→ See DoIt Alpha Pulse AI

Where Opus 4.7 consistently outperforms Gemini 3.1

This is the part I expected based on the public benchmarks, and the live runs confirmed it: Opus 4.7 is noticeably better at structured reasoning under ambiguity.

What that looks like in trading terms: When a setup is “almost there but not clean” — say, price is approaching a resistance, but momentum is weakening, and there’s news in 30 minutes — Opus does the patient thing. It returns “wait” with a proper explanation about why the asymmetry isn’t there yet. Gemini, in the same setup, sometimes commits. Not a hallucinated entry — a defensible one. But it’s the kind of trade that wins 50% of the time and loses you the other 50% by 1.2x. Over a month of those, you bleed slowly. Both models return a confidence score. The difference is what the score actually correlates with. Opus 4.7’s high-confidence trades tend to be the ones that pay off. When it says “0.85 confidence, structured continuation,” the trade hits target more often than not. When it says “0.55, marginal,” the result is closer to 50/50, exactly as a good model should self-report. Gemini’s confidence is flatter. A 0.85 from Gemini and a 0.65 from Gemini have less separation in actual outcome than the labels suggest. That’s not a fatal flaw — it just means you can’t use Gemini’s confidence as a position-sizing signal as cleanly as you can with Opus. This is small but it matters when you’re scaling cost: Opus tends to give you the verdict and a clean reason. Gemini sometimes returns three reasons, two of them redundant, when one would do. Over thousands of decisions, that adds up to a real API cost difference. If I told you Opus wins everything, this would be a marketing post. It isn’t. Gemini has clear advantages and you’d be wrong to dismiss it. Gemini 3.1 Pro returns a decision faster. Not “faster benchmark” faster — tradeably faster. On news-driven sessions where you have a 90-second window between a release and a setup forming, that latency difference matters. It’s also more consistent. Opus has occasional “thinking longer than usual” moments that, on a live trading account, can mean the setup is gone by the time the verdict arrives. Gemini doesn’t do that. If you’re running an EA that makes hundreds of decisions a month — including the “wait” decisions that don’t lead to trades — the bill matters. Gemini 3.1 Pro is cheaper per decision than Opus 4.7 by a meaningful margin. For traders who are testing on a small account, that’s the difference between a $30 monthly API bill and a $90 one. If you want the cheap-and-fast configuration, Gemini wins. If you want the careful-and-deep configuration, Opus wins. The good news is, on the same EA, you can switch on Sunday night and run a different model the next week. Where Opus shines on XAUUSD specifically, Gemini ports to other instruments slightly more cleanly. I’ve tested both on ETHUSD and XAGUSD — both via custom presets users sent in, by the way — and Gemini’s behavior changed less between instruments than Opus’s did. That isn’t a verdict. It’s a hint: if your strategy is “one model across many markets,” Gemini is the safer bet. If your strategy is “best decision per market, even if I have to retune,” Opus pays off on the markets it likes. Both models, with my prompt, struggle in roughly the same places. This is the part nobody writes about because it doesn’t sound good in a marketing pitch. If you’re shopping for an “AI trading bot” because you read that AI never loses, you’re going to be disappointed by reality. Both Opus and Gemini are tools that improve decision quality on average. They don’t make you immune to the market. No EA does that, AI or otherwise. Here’s the pattern I didn’t expect to see and that doesn’t show up in any public benchmark: The models disagree in ways that are themselves informative. When Opus says “take this trade” and Gemini says “wait,” in my data so far, the trade outcome is roughly even. Neither is reliably right. But when both say “take this trade” with high confidence, the win rate jumps significantly. And when both say “wait,” the setup almost never would have paid off if forced. That suggests a multi-model voting layer is more valuable than picking the “best” single model. It’s an obvious idea in hindsight; it’s also exactly the kind of thing you only discover when you’ve actually run two models on the same live account for a couple of months instead of writing a benchmark blog post. I’m building this voting logic into the next Alpha Pulse update. If you want to be on it when it ships, you already know where the newsletter is. Honest answer, based on real fills, real costs, real time-pressure: The actual answer most people don’t want to hear: it doesn’t matter as much as you think. The risk management around the AI matters more than the AI. A great model with bad position sizing loses money. A decent model with proper sizing makes money. Both Opus and Gemini are firmly in the “decent” range; what blows up accounts isn’t model choice, it’s risk discipline. Want to actually run this comparison yourself? DoIt Alpha Pulse AI is the EA that lets you bring your own API key for any provider — Opus, Gemini, GPT, Grok — and swap models on the panel without rebuilding anything.

Live Myfxbook · 179 verified trades · 4 community presets and counting

See Alpha Pulse AI →

Three lessons from this experiment, regardless of what you buy

Public benchmarks don’t transfer to live trading. The model that wins on coding tasks isn’t necessarily the one that wins on price decisions. Test what you actually care about, with money on the screen, before committing.
Cost-per-decision is the metric most people forget. A model that is 5% smarter but 3x more expensive is not a better trading model. Run the math on your decision frequency before you choose.
Multi-model voting beats single-model optimization. If you have the budget, running two models and taking only the trades they agree on is the highest-conviction setup I’ve found. The disagreement itself is information.

If you read this far and you’re still on the fence about whether AI-driven EAs are worth the effort: the honest answer is that they are, but not for the reason most YouTubers tell you. They don’t replace risk management. They don’t predict the market. What they do is make decisions that are slightly more contextual than a static rule-based EA could ever make. Over a year, “slightly more contextual” compounds.

FAQ

Is Opus 4.7 the same model people call “Claude Opus 4.7”?

Yes. Anthropic’s flagship reasoning model, released in 2026. The full name is Claude Opus 4.7. I use the short form throughout this post for readability.

Is Gemini 3.1 Pro the same as Gemini 3 Pro?

No. Gemini 3.1 Pro is the updated version released in 2026, more capable on structured reasoning and instruction-following than the original 3.0 series. If you’re testing AI EAs and your provider only supports 3.0, the experience will be noticeably weaker.

Can I really run both models on the same EA?

Yes, that’s the design of DoIt Alpha Pulse AI. You bring API keys for Anthropic, Google, OpenAI, xAI, DeepSeek or Qwen and switch in the panel. There’s no model lock-in. More setup questions answered here.

What about GPT-5.5 and Grok 4.20?

Both supported in Alpha Pulse. I tested them less rigorously in this round because Opus and Gemini are the two models my live forward test runs primarily on. GPT-5.5 sits between Opus and Gemini in my experience — slightly better than Gemini on reasoning, slightly worse than Opus on confidence calibration. Grok is wildly inconsistent in trading contexts; useful as a third opinion in a voting setup, not as a primary.

How much does running this cost in API fees?

Depends on the model. Gemini 3.1 Pro on a typical week: $15-25. Opus 4.7 the same week: $40-70. Running both for a voting layer roughly doubles. None of these costs include trading capital, slippage, or commission — those are separate.

Where can I see the live results?

The DoIt Alpha Pulse AI live forward test account is on Myfxbook, public. The link is on the product page under live results. Every winner, every loser, every drawdown period — visible without having to take my word for any of it.

Trading AI EAs involves real risk to capital. The performance described is from a live forward-test account on Axi Select / RoboForex. Past results don’t guarantee future ones. Run on demo first; size at low risk; treat any AI as a decision-support layer, not a guarantee.

Want updates when the multi-model voting layer ships? The newsletter is where it gets announced first: subscribe here.

Diego Arribas

Founder · DoItTrading

Building MT4/MT5 expert advisors and writing about prop-firm scaling since 2021. Currently running Alpha Pulse AI live on XAUUSD and trading Axi Select in parallel. I write what I'd want to read before paying for any of this myself.

All Insights → Trading Bots → Newsletter → Twitter →

Opus 4.7 vs Gemini 3.1: What Running Both on a Live Forex EA Actually Showed Me