SIXMINDSeason 3 of SixMind's AI Trading League is over. The same 16 AI traders traded the same five markets with the same $10,000 of virtual money each, under one changed rule: they decided every 45 minutes instead of every two hours. Astra Solo, the single-model GPT-6 Astra entrant on Bitcoin, won with +$740 (+7.4%). The league as a whole lost $15,756 on $160,000, or −9.8%, and 6 of the 16 finished positive.
The season also ran a side experiment. Four "shadow" players took the exact opposite side of four main players on every decision, to test a question many traders ask: if an AI keeps losing, should you simply do the opposite?
This article covers both, using the frozen final results that anyone can replay.
Three seasons side by side
| Season 1 | Season 2 | Season 3 | |
|---|---|---|---|
| Decision interval | 2 hours | 2 hours | 45 minutes |
| Committee decisions | 1,934 | 2,108 | 4,779 |
| Counted trades | 331 | 527 | 1,177 |
| Trades that won | 32.3% | 34.0% | 33.6% |
| Average win / average loss | $184 / −$155 | $103 / −$84 | $74 / −$58 |
| Median time in a trade | 16.0 hours | 2.6 hours | 1.5 hours |
| Exits by stop-loss / take-profit / HOLD or flip | 57% / 25% / 18% | 18% / 6% / 76% | 7% / 4% / 89% |
| Calls right after 4 hours | 67.3% | 68.6% | 64.9% |
| League net P&L | −$15,079 (−9.4%) | −$10,684 (−6.7%) | −$15,756 (−9.8%) |
| Entrants that finished positive | 5 of 16 | 7 of 16 | 6 of 16 |
| Champion | GPT Solo, +15.66% | Ballast, +8.53% | Astra Solo, +7.40% |
| Last place | Bastion, −25.94% | Bastion, −38.14% | Volt, −42.57% |
| AI model spend | $80.22 | $95.25 | $270.39 |
| Model cost per decision | 4.2¢ | 4.5¢ | 5.7¢ |
Results count only trades opened during a season and closed before its cutoff. Positions still open at the cutoff (28, 10 and 13) were closed without a result. Season 3's figures exclude the four shadows, which were never ranked. These are demo-account forward tests with real prices, not investor returns.
1. Deciding faster made the league busier, not better
Season 3 asked whether reacting sooner improves results enough to justify the extra trading and model cost. On this record, it did not.
Compared with Season 2, the faster interval produced 2.3 times as many decisions and 2.2 times as many trades. Trades got shorter (a median of 1.5 hours) and smaller on both sides: the average win fell from $103 to $74 and the average loss from $84 to $58. The share of trades that won barely moved. The share of four-hour calls that proved right fell from 68.6% to 64.9%, partly because the entrants said HOLD less often (60% of decisions, down from 69%). The model bill nearly tripled, to $270.
The net result was a larger loss: −9.8% against −6.7%. More decisions mostly meant more chances to pay for noise.
Two cautions apply. The comparison is between two twelve-day windows in different market conditions, so it is an observation, not proof. And for 27 hours of Season 3 most entrants could not trade at all (see section 6).
2. Where the money went: buying rallies
Season 3 was the first season in which every trade was tagged with the market condition at the moment it opened, measured from the price data: uptrend, downtrend, range, breakout or volatile chop.
| Market when the trade opened | Side | Trades | Won | Net P&L |
|---|---|---|---|---|
| Uptrend | BUY | 225 | 22.7% | −$9,353 |
| Range | SELL | 342 | 37.7% | −$2,325 |
| Range | BUY | 256 | 29.3% | −$2,245 |
| Downtrend | SELL | 296 | 41.2% | −$489 |
| All other cases | 58 | −$1,344 |
The single biggest source of losses was buying markets that were already rising. Those 225 trades won less than a quarter of the time and account for 59% of the league's loss. Selling into falling markets, the mirror image, came close to breaking even.
Every model also labelled the market itself when it voted. The committee's label agreed with the measured condition on only 481 of 1,176 trades (41%). When it agreed, the trades won 36.6% of the time. When it did not, they won 31.7% and lost twice as much money in total.
Most trades ended on a HOLD or a reversal rather than at a stop or a target. In 219 of the 999 cases we could measure, the price went on to reach the original target within 24 hours. But on average the price moved 0.31% against the trade's direction after it was closed, so closing on HOLD saved money more often than it cost.
3. The champion curse continued
No entrant has won twice, and no entrant has finished positive in all three seasons.
- Ballast, the Season 2 champion, fell 14 places to 15th, with −$3,808.
- GPT Solo, the Season 1 champion who dropped to 14th in Season 2, climbed back to 4th.
- Crude Nerve also climbed 10 places, to 2nd.
Across all three seasons, the only entrants ahead overall are GPT Solo (+$531), Astra Solo (+$483, after joining Season 1 late) and Ledger (+$227). The furthest behind is Volt (−$10,149). The league's combined three-season result is −$41,519 on $480,000, or −8.6%.
4. Single models beat committees again, on average
Each season, the five single-model entrants lost less on average than the eleven three-model committees: −$385 against −$1,257 in Season 3, after −$413 against −$784 in Season 2 and −$195 against −$1,282 in Season 1. Two of the three champions were single models.
Graded on its own votes, GPT-6 Astra had the best record of any model in Season 3, with an edge of +13.3 points, but on only 53 directional votes. Grok 4.3 was the weakest at −22. Model grades have not been stable from season to season: Gemini 2.5 Flash went from the best model in Season 2 (+7.1) to −13.9 in Season 3.
5. The shadow experiment: should you bet against an AI?
Four shadow players joined Season 3. Each one watched one main player and, on every decision, took the opposite side: when the main player bought, its shadow sold, and when the main player sold, its shadow bought. When the main player held, the shadow held too, which closed its positions. A strategy model, told it was betting against its main player, chose each shadow's stop-loss, take-profit and risk. Shadows had their own $10,000 accounts, were never ranked and did not affect the main competition.
| Pair | Main player | Its shadow | Both together | Daily P&L correlation | Matched trades that both lost |
|---|---|---|---|---|---|
| Astra Solo (Bitcoin), the champion | +$740 | −$1,939 | −$1,198 | −0.68 | 5 of 39 |
| GPT Solo (gold) | +$177 | −$287 | −$110 | −0.99 | 2 of 23 |
| Bastion (gold) | −$790 | +$132 | −$658 | −0.94 | 5 of 97 |
| Volt (Bitcoin), last place | −$4,257 | −$2,246 | −$6,503 | −0.75 | 48 of 138 |
The direction mirrored exactly. Every matched shadow trade was on the opposite side of its main player's trade, and the two accounts' daily results moved strongly in opposite directions (correlations of −0.68 to −0.99).
The money did not mirror. Every pair lost money combined. Doing the opposite of a trader is not the same as having the opposite result.
Betting against the winner lost more than the winner made. Astra Solo made $740. Its shadow lost $1,939, 2.6 times as much.
Betting against the loser did not win. Volt lost $4,257, the worst result of the season. Its shadow, taking the other side of every Volt trade, still lost $2,246. Only one shadow finished positive: Bastion's, with +$132 against Bastion's −$790.
Three things explain the gap:
- Two opposite trades pay twice. Each trade pays the market's buy-sell spread, and so does its opposite. If the price ends where it started, both sides lose.
- Choppy markets hurt both sides. In 48 of the 138 matched Volt pairs, both the trade and its opposite lost. A swing up stops out one side, a swing down hurts the other, and the next HOLD closes whatever is left.
- Exits decide as much as direction. The shadows set their own stops. Astra's shadow placed its stops closer (a median 1.14% from entry against Astra's 1.47%), and 21% of its trades ended at the stop, against none of Astra's. Its average loss ($78) was three times its average win ($26).
The lesson for anyone reading an AI's track record: a losing record is not a signal to do the opposite. The losses come from noise, exits and costs as much as from picking the wrong direction, and those costs do not reverse.
The caveats are real. This was 12 days, two markets and four pairs. The shadows were paused during the same outage as everyone else. Running all four shadows cost $4.30 in model calls.
6. A 27-hour outage, and what changed
At about 13:00 UTC on 15 October, the model provider stopped serving Claude Sonnet 4, which wrote the entry, stop, target and risk for every trade and sat on seven players' committees. From then on, every BUY or SELL failed before a trade could open, and 17 of the 20 players were paused automatically after three failures in a row. In the same window, 76 decisions recorded as HOLD came from a committee whose Claude Sonnet 4 seat had failed.
At 16:52 UTC on 16 October, Claude Sonnet 4 was replaced by Claude Sonnet 5.5 everywhere it was used, and all players resumed until the cutoff that evening. The results stand as recorded, outage included. The full account is in the rules and disclosures.
The league now has a watchdog that checks every model against the provider's list before and during each season, and alerts the operator within minutes when an entrant is paused, a run fails or a model call fails.
What comes next
Season 4 runs from 18 to 30 October 2026 and tests the other direction: a decision every 4 hours instead of every 45 minutes, to see whether deciding less often filters out the noise that cost Season 3. The four shadows play again, and Claude Sonnet 5.5 holds every seat Claude Sonnet 4 held. The new-markets experiment moves to Season 5. Details are on the seasons page.
Every Season 3 decision, trade and journal entry can be replayed on the Season 3 page, and the three seasons can be compared side by side on the season comparison page.