Editor's note (updated 2026-05-12): This editorial has been revised to reflect Season 1 scope refinement. XAUUSD and USDJPY were removed from the experiment universe; the numbers and references below reflect the updated 4-instrument scope (NAS100, US30, US500, EURUSD). Original publication date preserved.

Season Scorecard

The Week in Macro

Week 2 belonged to one story: Trump-Iran. The headline risk that had been simmering since the start of the season detonated on Friday April 24, and the entire tape spent the back half of the week pricing what happens when geopolitics overrides every other input. That is the regime the AI Trading Benchmark traded into.

The opening sessions were quiet by recent standards. Monday and Tuesday produced no completed trades from Claude — the model entered NAS100 long at 26,595.80 on Monday but the position remained open as the week ended, never resolving to TP or SL inside the window. GPT was quiet through Tuesday as well, then took two losses on Wednesday inside a single morning window as US30 and US500 both moved against the model.

The dollar story shifted mid-week. DXY firmed from the prior week's lows as risk-off positioning built into Friday. Yields backed up. The risk-on instruments that had carried Week 1 became unreliable — Wednesday's chop took out two GPT stops on the index side inside two hours, the kind of move that only happens when intraday correlation regimes fracture.

Then Friday. Trump-Iran tension dominated every screen. Equity indices opened soft, ripped higher mid-session on a denial headline, then sold off again into the close. Intraday whipsaws hit every major instrument. Retail traders got chopped — every social feed by Sunday was a montage of stopped-out longs and stopped-out shorts inside the same hour. The benchmark took two trades into that environment: Claude went short US500 and got stopped. GPT went long NAS100 and took the partial at TP1 for +$1,149.12 before the position rolled. Combined Friday damage across both models: +$209.12. That is a very small number — and a green one — for a day that catastrophically blew up an unknown number of retail accounts.

The macro frame for Week 2 is straightforward. Headline risk is now the dominant input. Setups that worked under the disinflation tape of Week 1 — long indices on dollar weakness — stopped working when the geopolitical bid replaced the macro bid. The models that survive this regime are the ones that take fewer entries, size them smaller, and accept that some weeks the right answer is a flat tape and a clean weekend.

About reported results. Each setup defines three take-profit targets (TP1, TP2, TP3), but the broker closes the full position at TP1 — so the realized R-multiple is always TP1's distance from entry when any TP is hit, and -1R on a stop. The dollar P&L shown in this editorial is the actual broker close at TP1 (or stop) for each trade. TP2 and TP3 are reported as informational levels: how far price ran after the broker had already exited.

Equity Curve

Apr 20 — Apr 24

Claude's Week

Claude was the better analyst and the worse executor this week. The model's read on EURUSD on Wednesday — short at 1.17105, exit at 1.1679, +1.67R analysis ride for +$1,606.38 — was the cleanest pre-trade thesis any model has produced on this benchmark. The dollar firmed into the European session, Claude was already positioned, and the trade scaled all three targets without a meaningful retracement. That is what conviction-on-conviction trading looks like when the macro tape cooperates.

The other two trades did not cooperate. Claude took a NAS100 long on Wednesday at 26,957 and got stopped out at 26,917.60 inside the first hour for -$1,014.80 — the same instrument the EURUSD same-day session was a pure chop on the index side. Then Friday: US500 short at 7,125 stopped at 7,133.90 for -$940. Two losses, both inside headline-driven sessions, neither a thesis failure. Both were regime failures — entries that would have worked in a normal tape and didn't survive an abnormal one.

The character signature is consistent with what Claude showed in Week 1. The model takes more trades than GPT (3 vs 3 this week, but Claude's trade-per-day distribution was more concentrated), commits when it has a read, and accepts the variance that comes with higher frequency. The Wednesday EURUSD trade is the model in its best form. The Wednesday and Friday stops are the model in its worst form, but the worst form did not turn into a blow-up because Claude held to fixed risk per trade — every loss was within ~1R of the model's $50,000 demo capital, and the cumulative drawdown across the two losing trades was -$1,954.80, less than 4% of starting capital.

Season-end-of-Week-2 position: $50,610.09, up +1.22% from the $50,000 starting balance. After two weeks and 10 total trades, Claude's win rate is 50.0% and net R is +0.90. The Wednesday EURUSD trade kept the season number close to flat in R terms despite the loss-heavy stretch through Wednesday and Friday.

The thing to watch in Week 3 is whether Claude maintains entry frequency or pulls back. The Friday US500 short was a thesis-correct setup in the wrong regime. A model that learns the regime shifted will trade less next week. A model that doesn't will repeat the same losses. Either response is informative — the benchmark is built to expose exactly this kind of regime-adaptation question, and Week 3 is when the answer starts to surface.

GPT's Week

GPT had its first in-scope trades of the season this week. The model's Week 1 ledger was empty after the scope refinement, so Week 2 is effectively the start of the head-to-head comparison on the ledger. The cadence: zero trades Monday, zero Tuesday, two losses Wednesday, zero Thursday, one win Friday. Three entries total. Two losses, one winner, net -$742.28 for the week.

Wednesday was where the week went sideways. GPT took two trades in a single morning session and lost both. US30 long at 49,563 stopped at 49,490 for -$1,101.40. US500 long at 7,129.60 stopped at 7,121.30 for -$790. Two trades, two stops, all inside the same risk-off window. The combined Wednesday damage was -$1,891.40 — the worst single day GPT has posted on this benchmark.

But GPT did not chase. Thursday produced zero trades. Friday produced exactly one entry: NAS100 long at 27,229.20, scaled at TP1 for +$1,149.12. That trade was the only green print on the Friday tape across both models. GPT was on the right side of the only direction that paid in the headline-driven session, and it took the partial rather than holding for higher TPs that the whipsaw would have erased.

The character signature emerging from GPT's first in-scope week: trades less frequently than Claude, accepts smaller targets, and is more willing to sit out a session entirely. Three trades for the week, one winner, two losses. Win rate 33.3%. Net R -0.98. Net P&L -$742.28. The Wednesday double-loss stretch was painful, but the Friday winner covered more than half of the damage and kept the week from running into deeper drawdown territory.

Season-end-of-Week-2 position: $49,257.72, down -1.48% from the $50,000 starting balance. After two weeks and 3 total trades, GPT's win rate is 33.3% and net R is -0.98. The model is behind Claude on every aggregate stat. It is also still in the simulation with no catastrophic days.

The thing to watch in Week 3 is whether GPT's selectivity is conviction or fear. Two trades on Wednesday and one on Friday is the right cadence for a chop-and-headline regime. But selectivity that becomes paralysis is worse than overtrading — every week without a winner compounds the season-return gap. The Friday partial proves the model still has the read; the test is whether it pulls the trigger at the right moments in Week 3's heavier macro calendar.

Trump-Iran Friday: Why the Whipsaw Mattered

The defining session of Week 2 was Friday April 24, and the defining input was geopolitics. By the open, Trump-Iran tension was the only headline that mattered. Equity futures gapped lower, ripped on a denial wire mid-morning, then sold the rip into a soft close. Bond yields backed up and then collapsed. Every cross-asset correlation that had held through Week 1 broke inside a single session.

Two trades hit the benchmark that day. Claude shorted US500 at 7,125 and got stopped at 7,133.90 for -$940. GPT went long NAS100 at 27,229.20 and scaled at TP1 for +$1,149.12. Combined day P&L across both models: +$209.12. For context, that is a green print on a session that, by every social-media account from Sunday, blew up an unknown number of leveraged retail traders, and the worst-case combined outcome was a small net positive.

That is the headline of Week 2. Not "Claude won" or "GPT won." Both models lost on net for the week — Claude -$348.42, GPT -$742.28. But neither model produced a single-day catastrophic loss, neither doubled down after a stop, and neither chased a reversal that would have erased a week of work in a single afternoon. Risk discipline is the season-level story.

Why Claude Outperformed on Wednesday

The single best trade of the week was Claude's EURUSD short on Wednesday April 23. Entry 1.17105, exit 1.1679, R-multiple +1.67, dollar return +$1,606.38 on the analysis ride. The dollar firmed into the European session as risk-off positioning built ahead of Thursday's calendar. Claude was short EURUSD before the move, scaled at TP1, held into TP3, and the trade closed without a meaningful retracement.

That is conviction-on-conviction trading. The thesis (dollar firming on shifting Fed expectations and risk-off flows) was correct. The execution (entry, sizing, target placement) was correct. And the timing — Wednesday, before the Friday headline shock — was correct in the sense that Claude captured the directional move before the regime fractured. After Friday's whipsaw, EURUSD reversed back through Claude's entry. A trader who held the position into the headline session would have given back most of the gain. Claude didn't, because the analysis-side targets resolved cleanly inside the window.

The same Wednesday session also produced Claude's NAS100 long at 26,957 — stopped at 26,917.60 for -$1,014.80. The model nailed the EURUSD setup and missed the index setup in the same morning. That is the variance any high-frequency analyst absorbs. The Wednesday net for Claude was +$591.58 across two trades, which is what a 50%-win-rate distribution looks like when one of the wins is a +1.67R outlier.

GPT's Disciplined Friday: The Lone Green Print

GPT's NAS100 long on Friday was the only winning trade either model took during the headline session. Entry 27,229.20, exit at TP1 27,321.90, +$1,149.12 on a +1.02R blended exit. The position scaled at the first take-profit and closed before the afternoon reversal. That is exactly the right execution for a headline-driven chop session — take the partial, lock the profit, and let the rest of the day produce zeros.

GPT did not take a second trade Friday. Did not chase the reversal, did not flip short on the rip, did not add to the winner. One entry, one partial, one closed position. Compare this to Claude's Friday — a single US500 short stopped inside the same whipsaw window — and the instrument-selection difference is the difference between green and red on the day. GPT's selectivity in chop is the model's most interesting Week-2 character trait.

The Wednesday double-loss complicates the GPT picture. Two stops in a single session, -$1,891.40 combined, is an ugly print regardless of context. But the model recovered most of it inside a single Friday partial, which suggests that the Wednesday losses were not thesis failures so much as concentrated risk in a single regime window. A model that takes two trades inside the same risk-off morning is not diversifying timing; it is concentrating it. That is a tractable mistake.

Same-Direction Overlap: Zero on Friday

Week 2 produced zero same-direction overlaps on the most important day. On Friday, Claude was short US500; GPT was long NAS100. Two different instruments, two different directions on the index side, no shared thesis.

That divergence matters because it tells you the models are reading the regime differently. In a clean directional tape, two AI models can agree on direction and split on instrument. In a chop tape, they can disagree on direction entirely — and on Friday, they did. Whether Claude's read or GPT's read was "right" is the wrong question. The tape was a coin flip. The right question is which model sized the coin flip correctly. GPT did. Claude did not.

What the Season Scorecard Says

After two weeks and 13 total trades (Claude 10, GPT 3), Claude leads on every aggregate stat. Win rate 50.0% vs 33.3%. Season net R +0.90 vs -0.98. Closing balance $50,610.09 vs $49,257.72. Net P&L gap of $1,352.37 between the two models, in Claude's favor.

But here is the season-level story that the scorecard understates: both models are still in the simulation. After two weeks of which the second was a textbook headline-driven chop, both AI models are within 2% of their starting balance. The benchmark has not yet seen a -10% week from either model, has not seen a single day worse than -4% from either model, and has not seen the kind of revenge-trading behavior that turns a -1% week into a -5% week. The season-level takeaway after Week 2 is not who's winning. It's that nobody has lost yet.

The Trade of the Week

Trade of the Week for the AI Trading Benchmark's second editorial cycle goes to Claude's EURUSD short on Wednesday April 23. The setup was textbook. The execution was clean. The dollar-return was the largest single-trade print across both models for the week (+$1,606.38), and the R-multiple (+1.67R) was the only number above 1.5R that the benchmark produced in five sessions.

The setup context: Wednesday opened with the dollar already firming. DXY had bid up off the Tuesday lows on a pre-positioned risk-off shift ahead of the Thursday and Friday calendars. EURUSD was rejecting 1.1730 from below, the prior week's range high, and the four-hour structure had rolled over. Claude entered short at 1.17105 — the precise inflection where the rejection pattern resolved into directional sell flow — and placed a stop above the swing high. Risk was sized at $499.20, exactly 1% of capital at the time of entry. Position size was 13.06 lots.

What followed was a near-perfect resolution. Price walked through TP1 inside the European session, scaled at TP2 in the New York morning, and closed the analysis-side at TP3 (1.1679) before the New York lunch chop. Total move from entry to TP3: 31.5 pips. Total time in trade: roughly six hours. No meaningful retracement against the position from entry to final exit.

The reason this is the Trade of the Week and not just a routine TP3 hit is the timing. Claude had this trade on the books before Friday's regime shift. EURUSD reversed back through 1.17105 on the Friday Trump-Iran headlines — a trader holding the position into Friday would have surrendered most of the gain. Claude exited cleanly inside the Wednesday session, captured the full analysis-side R, and was flat on EURUSD before the headline tape arrived. That is what timing-aware conviction looks like. The next sections show the broker execution and the pre-trade analysis Claude actually produced.

Account Performance

Profit taken at TP1 — the full position is closed at the first target to keep results measurable and comparable across models.

What to Watch in Week 3

Week 3 starts Monday April 27 and runs through Friday May 1. The macro calendar carries three pivot points: the FOMC decision on Wednesday, advance Q1 GDP on Thursday, and the April nonfarm payrolls on Friday. Each is potentially regime-defining on its own. Stacked into a single week — and on top of an unresolved Trump-Iran headline tape — the setup is the most consequential calendar of the season so far.

The question for both models is regime adaptation. Week 2 demonstrated that the macro frame from Week 1 (disinflation, dollar weakness) no longer holds. The new frame is headline-driven, correlation-broken, and intraday whipsawy. A model that recognizes that and trades less should outperform a model that runs the Week-1 playbook into a Week-2 tape. Through two weeks, GPT has been more selective; Claude has been more active. Whether that posture inverts in Week 3 will tell us how each model is reading the regime.

Three specific watchpoints: First, does Claude continue to take indices long into Friday sessions, or does the Week-2 result cause a pullback? Second, does GPT take a trade Wednesday into the FOMC, or sit it out and trade only the post-decision tape? Third, do the models converge on a same-direction trade — and if they do, what is the instrument? Same-direction overlap has been absent across the first two weeks.

The benchmark itself remains intact. Two weeks down, both models within 2% of starting capital, no blow-ups, and a season scorecard that is actually getting interesting. Week 3 either separates the models or keeps the gap tight. Either outcome is reportable.

Methodology

This weekly editorial aggregates trading results from April 20-24, 2026. All numbers come from the live broker execution ledger — no simulation, no backtest.

How P&L is computed. Week P&L is calculated as weekEndBalance - weekStartBalance, never as the sum of individual trade net P&L. The two can differ slightly due to rounding in partial exits; the broker balance is always authoritative.

Week rollover. Each week's starting balance is the previous week's ending balance. Week 1 uses the experiment's initial capital ($50,000 per model). This is why account balances — not trade sums — are the ground truth for performance tracking.

Net R vs. Net P&L. Net R is a risk-adjusted measure (sum of each trade's reward/risk multiple). Net P&L is the literal dollar change in account balance. Both are reported; R-multiples are more comparable across instruments with different tick values.

Weekend handling. Daily balance series forward-fill Saturday and Sunday from the prior Friday close, since markets are closed. This keeps chart visuals continuous without fabricating activity.

Methodology stability. Rules don't change mid-phase. If any rule is updated for a future phase, it's documented at the methodology page.

Scope refinement. This editorial was retroactively updated on 2026-05-12 to remove XAUUSD and USDJPY from the experiment universe.