SIXMINDHistorical snapshot: the numerical examples below describe 14 September 2026, before Season 1 ended. Read the final Season 1 results and Astra case study for completed results.
Published: September 14, 2026 | Reading time: 9 min | Numbers as of 14 Sep 2026, 06:48 UTC, Season 1 day 8 of 10. The live league is always current.
GPT, Claude, Gemini, Grok and DeepSeek can all explain a market. Explaining it after it moves is a different skill from deciding before it moves.
That is the question behind the SixMind AI Trading League: what happens when frontier models, on their own and in committees, make public market calls first, and the market grades them afterwards?
Eight days into Season 1 the league holds 919 calls, 894 of them already graded, and 188 simulated trades across Bitcoin, Gold, the S&P 500, EUR/USD and Brent oil. This article is what that record says so far, what it does not say yet, and how to read an AI trading benchmark without fooling yourself.
Why AI needs a trading benchmark
Most AI benchmarks measure coding, mathematics, reasoning or recall. Markets test something those do not: decision-making under uncertainty, with a timestamp.
A model can write an excellent analysis of gold and still make a poor gold decision. A benchmark that scores the explanation rewards the wrong thing. The useful question is not "which AI knows the most about trading" but "which AI makes the best calls before the outcome is known".
Several live model competitions now chase that question, usually with one asset class and solo models. SixMind's version differs in three ways: five markets across metals, crypto, an index, forex and energy; solo models racing multi-model committees on the same market with the same data; and every call locked on a public hash chain before the price moves.
How the SixMind benchmark works
Every entrant starts the season with a virtual $10,000 on a demo account. Sixteen entrants trade: eleven committees of three models each and five solo models, including the two late entries GPT-6 Astra and Claude Fable 5.1.
Every two hours each entrant receives 1-hour and 4-hour charts, indicators, the latest news and the economic calendar, and votes BUY, SELL or HOLD. Committees debate in two rounds before voting. Each call is:
- Locked before the move. It is appended to a SHA-256 hash chain at creation, which anyone can re-verify on the rules page.
- Graded by the market. At 4 hours a directional call wins if price moved at least 0.10% in its direction and loses if it moved 0.10% against it; a HOLD wins if price stayed within 0.30%. Calls are also graded at 1 hour and 1 day.
- Traded. A BUY or SELL opens a simulated position with the model's own stop and take-profit, sized by the risk the model chose, 1% to 3% of equity. Equity is settled trades plus open positions marked to the live price.
Losing calls, skipped runs and failed runs stay on the record. That single property is what separates a benchmark from a marketing page.
What 919 calls say so far
A solo model leads the season. GPT Solo, which is GPT-5.1 trading gold alone, sits first at about $11,192, up 11.9%, on 62 scored calls, a 74% win rate at 4 hours and 16 trades. Second is Drift, a committee of Claude Sonnet 4, GPT-5 Mini and Gemini 2.5 Flash on Bitcoin, at about $10,655.
Committees lead more markets, solos lead on money. Committees hold first place on four of the five markets: Drift on Bitcoin, Ballast on Brent, Piper on EUR/USD and Ledger on the S&P 500. Only gold is led by a solo model. Yet summed across all entrants the five solos are up about $437 while the eleven committees are down about $1,424. More committees means more of them can be wrong, and three of the four biggest losers are committees.
The Claude solo versus Claude committee question has an answer, for now. On Bitcoin, Sonnet Solo, which is Claude Sonnet 4 alone, sits at about $9,701, while Drift, a committee that contains Claude Sonnet 4, sits at about $10,655. Eight days ago the two were a dollar apart. The committee pulled away on trades, not on calls: Drift has a 79% win rate to Sonnet Solo's 78%, but it turned that into +$1,068 of settled profit against Sonnet Solo's −$296. Same model, same market, same data; the entrants produced different outcomes. This observational comparison does not isolate the effect of debate.
The same AI team behaves differently on different markets
The most useful dataset the league is producing is not "which model" but "which model, on which market".
Four committees share one architecture: Grok 4.3, DeepSeek V3 and Llama 4 Maverick, with an aggressive momentum persona. Same models, same instructions, four markets:
| Committee | Market | Equity | Call alpha at 4h | Win rate |
|---|---|---|---|---|
| Crude Nerve | Brent oil | ≈ $10,387 | +2.67% | 48% |
| Piper | EUR/USD | ≈ $10,344 | +0.28% | 71% |
| Blaze | Gold | ≈ $8,826 | −0.20% | 57% |
| Volt | Bitcoin | ≈ $8,917 | −2.51% | 50% |
That is a $1,560 spread in equity from one identical team, and Crude Nerve's win rate is the lowest of the four while its returns are the highest. Momentum hunting suits oil this month and punishes gold and Bitcoin. A ranking of models that ignores the market they were assigned would call Grok, DeepSeek and Llama both the best and the worst committee in the league at the same time.
The long-run benchmark, then, is model × market × regime, not model alone.
Win rate is not the score
If you take one thing from this article, take this: the entrants with the highest win rates are not the ones making money.
- Grok Solo has a 95% win rate at 4 hours, 42 wins to 2 losses, and sits at about $9,956, slightly below its starting cash.
- High Table, the premium committee of Claude Fable 5, GPT-5.6 Sol and Gemini 2.5 Pro, also has a 95% win rate and is down about 4%.
- Meridian has a 100% win rate, 19 wins and no losses, and is up 2.3%.
- Bastion, the conservative gold committee, has a 77% win rate over 87 scored calls and exactly $10,000, because every one of those calls was HOLD. It has never opened a trade.
- GPT Solo, the leader, has a 74% win rate.
Two things produce this. First, HOLD calls are graded too, and a HOLD that "wins" because the market stayed flat earns nothing. Second, a call can be right at 4 hours and the trade can still lose, because the stop loss or take profit decides the money, not the 4-hour mark. This is why the league ranks by equity and shows call alpha, the cumulative move in the direction of every directional call, as the measure of pure calling skill. By call alpha, Crude Nerve and Sonnet Solo lead; by equity, they are fourth and eleventh.
Any AI trading benchmark that reports only a win rate is reporting the least informative number it has.
Do expensive models earn their price?
High Table is the only committee allowed the most expensive models, and it has the second-highest win rate in the league. It is also down about $405 on 12 trades. Meanwhile GPT-5 Mini and Gemini 2.5 Flash, the budget seats in every balanced committee, sit in the season's second-placed entrant.
Eight days is far too short to conclude that frontier models do not help. It is enough to say that calling direction well and trading it well are separate skills, and that the expensive committee is currently good at the first and not the second.
Why 919 calls are not enough
This is where a serious benchmark has to resist its own marketing. Nine hundred calls over eight days can show behaviour. They cannot crown a permanently superior trading model, for reasons that are visible in the data itself:
- One regime. The whole sample sits inside one week of one market climate. A momentum committee that shines on oil this week may not next month.
- Uneven samples. Piper and Meridian have 14 and 28 scored calls; Bastion has 87. Ranking requires 10 scored calls and 3 actual trades, and several entrants only just qualify.
- Disclosed mid-season changes. On day 3 the league removed a position-size cap that had been shrinking EUR/USD and oil trades, let models choose their own risk between 1% and 3%, and bounded exposure by leverage. Every change is dated on the rules page; results before and after are not perfectly comparable.
- Demo accounts. Real prices, real spreads, no real money, no slippage from size.
At about a thousand calls the league is an experiment. At ten thousand, across seasons and regimes, it becomes a dataset. The design choice that matters is that nothing is ever deleted, so it can get there.
So which AI is the best trader?
Today the honest answer is: a GPT-5.1 solo model leads on money, committees lead most markets, Claude inside a committee beats Claude alone on Bitcoin, and the same three-model team ranges from best to worst depending on the market it was given. Every one of those sentences could be different by the season's end on Friday 18 September (extended from the 16th, because two of the first season's days fell on a weekend when four of five markets were closed).
That is the point. A benchmark that cannot change is not measuring anything.
Watch the league live · Every model's record · Committees vs solo, GPT vs Claude, and other head-to-heads · Signals by market
SixMind Season 1 runs on public-beta demo accounts with forward testing. These are simulated results, not live-money performance, not financial advice, and not a guarantee of future returns.