We gave five AI models $100,000 each. One never placed a single trade.

This is a live, ongoing experiment, not a backtest. Every position and every stated rationale is public on the AI vs AI report. The numbers below are a snapshot; the live ones will have moved.

The standings, 31 August 2026

Each model started with $100,000 in virtual cash and picks real stocks at real prices. Each must write a rationale for every trade.

#	Model	Value	Return	Positions	Cash	Largest position	Last trade
1	GPT-4o (OpenAI)	$115,041	+15.04%	6	1.8%	MSFT — 39.1% of book	3 days ago
2	Llama 3.3 70B (Meta/Groq)	$104,734	+4.73%	8	0.0%	XOM — 20.8%	55 days ago — model dead
3	Claude Sonnet 4.5 (Anthropic)	$103,132	+3.13%	8	4.8%	VOO — 58.6%	32 days ago
4	Grok 4 (xAI)	$100,000	0.00%	0	100%	—	never
5	Gemini 2.5 Flash (Google)	$98,588	−1.41%	1	50.7%	QQQ — 49.3%	70 days ago

Four facts jump out before you get anywhere near "which AI is smartest."

1. One model never started

Grok 4 holds zero positions and $100,000 in cash. Not a conservative allocation — literally no trades, ever, while the other four have been trading for months.

We checked whether this was a malfunction. It isn't. Given the exact prompt and the exact 100%-cash position it faces every weekday, Grok returns well-formed JSON:

{"decisions": []}

The API responds in under a second, the output parses cleanly, and the answer is no trades. Every time, for months.

This is worth sitting with, because it's the failure mode nobody writes about. The interesting question isn't whether an AI can pick a winning stock; it's whether it will act at all under uncertainty. Given an open-ended mandate and no forcing function, one of five frontier models simply doesn't.

Update, later the same day. Having established that the silence was a choice rather than a fault, we changed the rules for everyone: models must now justify a decision to hold in a public hold_reason, the way they already justify a trade. Grok began trading immediately on the very next run. That tells you something sharper than the original finding did — the inaction wasn't conviction, it was the absence of a requirement to explain itself. Asking a model to account for doing nothing was enough to make it do something. Make of that what you will before your next portfolio review.

The original prompt ended with the line "Passing is a legitimate choice. Don't force trades." We meant it as a guard against churn. One model read it as permission to never begin.

Human investors do this constantly. It's called waiting for a better entry point, and it's the most common way people spend a decade in cash.

2. One was dead and we couldn't tell

Llama last traded on 7 July. We assumed it was holding. It wasn't — it was broken.

Groq deprecated llama-3.3-70b-versatile on 17 June 2026 and stopped serving it in August. Every call the tournament made returned:

404 — The model 'llama-3.3-70b-versatile' does not exist or you do not have access to it.

The reason nobody noticed for seven weeks is the part worth writing down. Our scheduler caught the exception, logged it to stdout, and returned an empty decision list — which is byte-identical to a model deciding not to trade today. The daily job cheerfully reported "ok, 0 trades" either way. A dead competitor and a patient one looked exactly the same from the outside.

That's a general lesson about automated systems, not just this one: if a failure and a legitimate no-op produce the same output, you will not find the failure. We've since changed the job to report provider errors as failures and shipped a probe script that calls every model directly.

Llama's $104,734 is therefore a frozen snapshot from 7 July, not a live result. It has been replaced in the tournament by Qwen 3.8 27B (Alibaba), starting fresh at $100,000 — Groq no longer hosts any general-purpose Meta Llama model, so there was nothing to swap in on the same key.

Gemini, by contrast, is genuinely quiet: we probed it and it responds normally, it has simply chosen not to trade since 22 June.

3. The leader is winning with the riskiest book

GPT-4o is first, and it got there by being the most concentrated and least hedged: 6 positions, 1.8% cash, and 39.1% of the entire portfolio in a single name (Microsoft).

If a user did that in Sydnical, our coach would mark the position sizing a Mistake — not because it lost money, but because it's a bet whose outcome is dominated by one company. It happens to be working. That's precisely why grading on outcomes is a trap: a bad process producing a good result is the single most dangerous feedback an investor can receive, because it gets repeated with more money.

Ask the honest counterfactual. Had Microsoft dropped 30%, GPT would sit near the bottom of this table with the same reasoning, unchanged.

4. The "most cautious" model partly opted out

Claude Sonnet 4.5's largest holding is VOO — an S&P 500 index fund — at 58.6% of the book. In an experiment about stock picking, the most methodical participant put the majority of its portfolio into not picking stocks.

That's arguably the most defensible allocation in the table. It's also, by the standards of the game, a refusal to play. Both readings are true, and the tension between them is the whole point.

What this does and doesn't tell you

It does not tell you which AI is the better investor. Ten weeks is noise. Barber and Odean's work on day traders needed 15 years of Taiwanese exchange data — 3.7 billion transactions — before it could separate skill from luck, and even then found only about 1% of day traders were predictably profitable net of fees. A five-participant, ten-week sample can't distinguish anything.

It does tell you something more useful, because behavior is visible immediately even when skill isn't. In ten weeks we can already observe: who acts, who stalls, who concentrates, who diversifies, who keeps a thesis, and who quietly stops showing up. Those traits are stable and they're measurable long before returns mean anything.

Which is the same reason we grade human users on decisions rather than on their balance.

The uncomfortable parallel

Every one of these models writes excellent rationales. Fluent, specific, plausible. Here is GPT's most recent, on adding to Amazon:

"Amazon's current position in e-commerce and cloud services, combined with a steady increase in consumer discretionary spending, positions it well for continued growth."

That's a perfectly reasonable sentence. It's also indistinguishable in quality from a rationale that would precede a terrible trade — and that's the lesson. A convincing reason is not evidence of a good decision. The models produce beautiful arguments for concentrated bets, for sitting in cash, for buying an index instead of doing the work. So does every confident voice in your feed. So, probably, do you.

The only way to find out whether your reasoning holds up is to make real decisions and grade the behavior — which is exactly the lens we point at these five models, pointed back at you.

Methodology

Five models, one paper portfolio each, $100,000 starting virtual cash, real market prices, no leverage and no derivatives. Every trade requires a written rationale, stored and published. No model sees another's book. Models are prompted on a schedule; none is forced to trade, which is why "never traded" is a possible and observed outcome. Figures above are as of 31 August 2026 and will drift — the live report is authoritative.

This is educational research, not investment advice, and nothing here is a recommendation to buy or sell anything.

FAQ

Which AI model is best at picking stocks?

On this experiment's data as of 31 August 2026, GPT-4o leads at +15.04% and Gemini 2.5 Flash trails at −1.41%. But a ten-week window with five participants cannot separate skill from luck — academic work on trader skill has required over a decade of data to do that. Treat the ranking as an observation about behavior, not a verdict on ability.

Can AI beat the stock market?

No large language model has demonstrated reliable price prediction, and short-run results are dominated by chance. What this experiment shows is that models differ enormously in behavior — one has never traded at all, two stopped after a few weeks, and the current leader is also the most dangerously concentrated.

Why did Grok never make a trade?

Not because it's broken. We probed the model directly with the tournament's own prompt: it responds in under a second with valid JSON containing an empty decision list, {"decisions": []}. Given an open-ended mandate and no requirement to act, Grok simply chooses not to — every day, for months. This mirrors a very common human failure mode: waiting indefinitely for a better entry point.

Is the AI vs AI experiment real money?

No. Each model runs a $100,000 paper portfolio at real market prices with no real capital at risk. Every holding and every written rationale is public so the reasoning can be judged, not just the returns.

How can I trade against the AI models?

You can run your own portfolio with the same live prices for free and have every decision graded by the AI coach. Follow the standings on the live report, then start your own book.

---

The scoreboard will keep moving. The behavior underneath it — who acts, who concentrates, who quietly stops — is the part worth watching, and it's the part that predicts what happens next. See how your own decisions grade →