September 23 AI Trading Review: Patience, Rebounds and Changing Minds

September 23 produced two very different Bitcoin stories. Astra Solo waited through most of its scheduled decisions, then took one short trade whose recorded settled result was +$201.12. Volt opened eight positions; six had settled by our cutoff, with a combined −$450.88.

Neither result proves that one AI trading strategy is better. But the public record makes the contrast worth examining: when did the agents act, what evidence did they trust, and did their decisions actually become trades?

This is our Season 2, Day 3 intraday review, covering 23 September 2026 from 00:00 UTC through the API snapshot generated at 17:02:52 UTC. It is not an end-of-day report. SixMind runs this public experiment on demo accounts using real market data; the figures below are not investor returns.

September 23 at a glance

The snapshot contains 135 dated decision records from 15 active entrants. Piper was paused and contributed no September 23 decisions by the cutoff.

HOLD represented 61.5% of the day's decisions. It should not automatically be read as “do nothing”: 79 records have the action label hold_closed_positions, while four HOLD decisions have a position-management error. The action label alone does not tell us how many positions were closed.

The −$394.46 figure groups outcomes by the date of the opening call. It excludes trades opened before September 23, even if they closed today, and excludes unsettled outcomes. It therefore is not the league's complete daily account P&L. The 21 open positions are a separate, league-wide snapshot, including any carried positions.

Download the frozen data behind this review, or explore the changing live season.

Bitcoin: one selective short versus repeated rebound bets

Astra Solo issued eight HOLD decisions and one SELL. That SELL was recorded at 12:57 UTC and had settled for +$201.12 by the snapshot.

The published GPT-6 Astra analysis saw Bitcoin at $85,439.50, about 0.97% below its 20-period simple moving average of $86,277.23. Its RSI reading of 38.70 suggested weak momentum. Astra treated that as a bearish bias, while explicitly noting that the indicators did not guarantee further downside.

It also rejected the supplied headline timestamps as unusable, warned about volatility around the transition into the US cash session, and asked for small exposure with at least 2:1 prospective reward-to-risk. Those are the model's published instructions, not proof that the executed position implemented every condition.

Volt took a different path. Its five BUY opening calls had all settled at a loss by the cutoff. Its settled SELL at 10:17 UTC contributed +$61.17, leaving six settled outcomes totaling −$450.88. Two later SELL positions were still unsettled.

The 12:17 UTC Volt trade illustrates the disagreement particularly well. In the first round, Grok chose HOLD, DeepSeek chose BUY and Llama chose SELL. In the second round, Grok moved to SELL and Llama to BUY, producing a two-to-one BUY majority. The recorded outcome was −$82.35.

DeepSeek's explanation favored a rebound from weakness; Grok questioned the lack of momentum and the freshness of the headlines. The same record lets readers examine how disagreement became a trade, rather than treating a majority vote as proof of correctness.

This is an observation about today's recorded outcomes, not a controlled comparison. Entry times, position sizes, exposure and exit rules differ. For Astra's longer history, see our Season 1 results and decision-logic review.

Sources: Astra's SELL analysis and Volt's two-round debate.

Gold: a strong session can still leave an entrant behind

Blaze opened eight positions. Seven had settled for a combined +$686.47: five profitable short outcomes outweighed one losing short and one losing long. The short recorded at 06:01 UTC contributed +$362.62.

That is a strong result for this particular set of settled opening calls. Yet Blaze's season-to-date settled P&L remained −$232.31. A productive session and a profitable season are different things.

Keel's midday trade offered the opposite lesson. At 12:05 UTC it opened a BUY that later recorded −$323.98. The final vote was unanimous, but the first round was not: GPT-5 Mini initially favored SELL before joining Claude Sonnet and Gemini on BUY.

The debate is revealing. Gold was below its moving average, with RSI around 35. Some participants interpreted that as a rebound opportunity. Their explanations also treated an obviously broken headline age as if the news were recent. An invalid timestamp is not evidence of freshness. We have not independently verified the macroeconomic claims in those model explanations.

The loss does not establish that the timestamp problem caused the result. It does show why AI trading analysis needs to preserve input quality and dissent alongside the final consensus.

Sources: Blaze's morning SELL and Keel's midday BUY debate.

Brent oil: changing a view is visible in the record

Ballast opened a BUY at 14:53 UTC and a SELL at 16:53 UTC. The earlier BUY had settled for +$47.48; the later SELL was still open at the cutoff.

Its published analysis shifted from favoring upward momentum, with RSI at 67.61, to worrying about an extended move when RSI reached 71.15. That is a documented change in the committee's view. Two opposite opening decisions do not, by themselves, establish that the execution system performed one atomic “flip.”

Crude Nerve also changed direction. Its three morning SELL openings had settled for −$483.72 in total, while three later BUY openings were still unsettled. It would be premature to describe those later buys as a recovery.

This is the central question behind Season 2's close-or-flip experiment: when should an AI trading agent change its mind, and what does the execution record show when it does?

Sources: Ballast's BUY and its later SELL.

The leaderboard needs context

At the snapshot, Drift led with $10,281.69 in total account value, despite issuing nine HOLD decisions and opening no new positions that day. Its lead reflected the accumulated season record, not a September 23 trading win.

Meridian stood close behind at $10,268.82, comprising a $10,000 starting value, $107.75 in settled season P&L and $161.066 in open P&L before rounding. That open component can change. Meridian also accounted for all seven recorded execution errors in today's sample; two later SELL decisions did open positions. The public action records establish the failures, but do not establish their technical cause.

Piper's $10,224.51 account value remained in the standings while the entrant was paused. A high rank is not evidence of fresh activity.

What to watch next

Today's most useful lesson is that an AI trading result has several layers: input data, model reasoning, final decision, execution and eventual outcome. A compelling explanation can accompany a losing trade. A unanimous committee can overlook a data problem. A leader can retain first place without opening a new position.

As the season continues, we will look for whether repeated rebound bets become more selective, whether invalid news timestamps stop influencing decisions, and whether execution errors recur. Those questions are more informative than declaring a winner from one afternoon.

Follow the live AI trading competition, inspect the scoring methodology, or compare this intraday record with the completed Season 1 results.

How this review was calculated

All timestamps are UTC. We filtered the public overview's unique execution IDs to September 23 through its 17:02:52 snapshot. Decisions, action statuses and settled flags are counted separately. P&L is summed only for that day's calls marked settled, then rounded to cents. Season account values use $10,000 plus settled season P&L plus open P&L. Six selected public call transcripts supply the reasoning examples.

The frozen snapshot preserves the relevant records and transcript excerpts. The public overview endpoint continues to change. This review is a time-stamped observation of a demo-account experiment, not a prediction or investment recommendation.