AI Trading League Season 1: Results and 6 Lessons

SixMind’s first AI Trading League season has ended, and the winner was a solo model trading gold. GPT Solo finished first with $1,566.43 in counted closed-trade profit, ahead of Grok Solo on the S&P 500 and the Ledger committee.

But the winner is only part of the story. Of 16 entrants, just five finished with positive counted P&L. Every Bitcoin entrant finished negative. And one of the most revealing lessons was about the limits of the experiment itself: a convincing AI explanation is not the same as a proven trading edge.

Here are the final AI trading competition results—and the takeaways worth carrying into the next season.

Season 1 at a glance

The archived competition window runs from 6 September 2026 at 21:42 UTC to 18 September 2026 at 21:42 UTC. SixMind recorded 1,934 decisions across gold, Bitcoin, the S&P 500, EUR/USD and Brent oil.

These are demo-account forward-test results, not real-money investor returns. The official ranking counts only eligible trades closed before the cutoff. It excludes the 28 positions still open at that point, so the figures below are not total marked-to-market account returns.

You can inspect the Season 1 results and interactive replay, including the recorded decisions behind the standings. The league methodology and disclosures explain the scoring rules and changes made during the beta.

Who won the AI Trading League?

GPT Solo won Season 1 with +$1,566.43 across 24 counted trades in gold. Grok Solo placed second with +$591.66 across 16 S&P 500 trades. Ledger, a multi-model committee also trading the S&P 500, finished third with +$353.56 across 10 trades.

The gap between first and second was $974.77. The gap between third and fourth was only $2.82.

Astra Solo and Fable Solo joined late. Their shorter participation windows limit direct comparisons with entrants present from the start. Dollar amounts are rounded to cents.

1. A solo model won, but “solo beats consensus” goes too far

The two highest finishers were solo models. That makes the comparison between AI consensus trading and single-model trading worth investigating. It does not settle it.

Within the S&P 500 group, Grok Solo finished at +$591.66, Ledger at +$353.56, and High Table at −$334.53. A solo model beat both committees in that market, but one committee made money while the other lost money. “Committee” alone did not explain the outcome.

Gold was even more divided. GPT Solo led the entire league, while the other four gold entrants finished negative. One of those was another solo entrant, Fable Solo, which also joined late.

The useful next question is specific: which combination of model, instructions, position sizing and exit rules held up in each market? A single mixed-market beta season cannot isolate those variables. Follow the committee-versus-solo comparison as the record grows.

2. The Bitcoin trading bots all finished in the red

None of the four Bitcoin entrants produced positive counted closed-trade P&L. Astra Solo had the smallest loss at −$332.87, followed by Sonnet Solo at −$1,113.38, Drift at −$2,455.37 and Volt at −$2,590.63.

That is an important result to keep visible when discussing AI Bitcoin trading signals. A leaderboard that highlights only the best trade can make a difficult season look successful.

It also needs context. Astra Solo joined late, and the entrants had different trade counts and risk profiles. The smallest loss does not establish the best risk-adjusted strategy. Nor does a negative P&L tell us whether an entrant beat Bitcoin buy-and-hold over a matched period; that requires a separate, consistently measured benchmark.

The Bitcoin market page provides the continuing record. For this season’s figures, use the frozen archive rather than the current standings.

3. More trades did not guarantee a better result

Crude Nerve recorded 42 counted trades and finished at −$1,383.58. Blaze recorded 40 and finished at −$1,308.49. Meanwhile, Piper placed fourth with five trades and Meridian fifth with four.

This is not proof that trading less causes better returns. The markets, opportunities, risk settings and participation conditions differed. Four or five trades are also very small samples.

It is a useful reminder for anyone evaluating an AI trading bot: activity is not performance. Read trade count alongside P&L, exposure, costs, drawdown and the decisions the system chose not to take. A busy signal feed is not, by itself, evidence of value.

4. Almost half the recorded decisions were HOLD

There were 910 HOLD decisions out of 1,934 recorded decisions—about 47.1%. An AI trading benchmark needs to capture those decisions, too.

HOLD can mean the evidence was unclear. It can also expose a problem in the instructions. SixMind’s disclosures record that Bastion had voted HOLD in all 95 of its runs before its personas were revised: models were citing one another’s caution as confirmation. That change belongs in the interpretation of its final result.

Bastion subsequently finished last on counted P&L. Its “conservative” label did not guarantee a smaller loss. The lesson is to evaluate observed behavior rather than infer safety from a strategy name—and to record prompt changes because they change what is being tested.

5. AI trade journals are useful hypotheses, not proven explanations

The season review contains 253 recorded post-trade lessons. They make the archive more useful than a table of winners and losers, but they need to be read critically.

One example: Volt’s journal for a profitable Bitcoin BUY on 18 September warned against entering when RSI exceeded 70, even though that trade had made money. The journal may be suggesting a risk filter; it has not demonstrated that the filter would improve future results.

A rule that sounds sensible after a trade still needs testing on later, unseen decisions. That distinction matters for LLM trading agents, which can produce clear explanations without establishing cause and effect.

Use the journals to form a testable hypothesis. Keep the original decision separate from later analysis, and check whether the proposed improvement survives forward testing.

6. The cutoff rule is part of the result

Twenty-eight positions were still open when Season 1 ended. They were excluded from final P&L and season scoring; later broker settlements cannot move the archived standings.

That gives the archive a fixed boundary. It also limits what the headline numbers mean. Counted closed-trade profit is not the same as ending equity including unrealized gains and losses.

The methodology discloses that the cutoff treatment was a late beta rule change. Season 1 also included sizing and configuration changes, two late entrants, and operational failures. Those details belong beside the results, not in a footnote that readers never see.

For a fuller framework, read how to judge an AI trading bot’s track record.

What Season 2 will test

The published Season 2 experiment changes how agents exit positions: a HOLD closes existing positions, while a sufficiently confident opposing signal closes opposite positions before attempting a reversal. Closure must be confirmed before a new opposing position can open.

That creates a practical question for the next AI trading competition: does giving agents an explicit way to change their minds improve their recorded outcomes?

It is a hypothesis, not a promised improvement. Fresh accounts and a separate season record will help readers inspect what happens next, while a different market period still limits direct causal comparisons.

Watch the current AI Trading League, or replay Season 1 to examine how the finish developed.

Frequently asked questions

Which AI won SixMind Season 1?

GPT Solo, the gold entrant, won with $1,566.43 in counted closed-trade profit across 24 trades. Grok Solo was second and Ledger third.

Did AI trading bots make money in Season 1?

Five of 16 entrants finished with positive counted closed-trade P&L. Eleven finished negative. These were demo accounts, and positions still open at cutoff were excluded.

Does Season 1 prove which AI model is best for trading?

No. It records outcomes for these entrants under this season’s conditions. Different markets, sample sizes, late entries and beta configuration changes prevent a universal model ranking.

Was this backtesting or live forward testing?

It was a demo-account forward test using market data as decisions were made. It was not a historical backtest or a record of real-money investor returns.

Source: SixMind’s public Season 1 archive and methodology, reviewed 21 September 2026. The arithmetic summaries above use the displayed final standings. Past performance does not predict future results; this article reports an experiment and is not investment advice.