From GPT-5 launch to Claude consolidation: SkyAnalyst's 2026 year-to-date

Restated: Gold (XAUUSD) was part of SkyAnalyst's coverage from inception (Jan 12, 2026) through May 2026. We've since narrowed coverage to six instruments — EURUSD, GBPUSD, USDJPY, US30, NAS100, US500 — and these numbers are restated for the current lineup. The original publish date is preserved; cumulative figures have been recomputed.

SkyAnalyst went live on January 12, 2026 with GPT-5 (c1/openai/gpt-5/v-20250930) as the master analyst. Through May 13, the published record carries 92 master-automation trades, 53 winners and 39 losers, a 57.6 percent win rate, and plus 15.62R net on the TP1 baseline. A simulated $100,000 account at 2 percent risk per trade sits at $131,242.81 on the static line, or $134,086.32 compounded, four months and a week into the live record. The year had a shape, and the shape had three execution eras under one architecture. GPT-5 ran the launch from Jan 12 through Mar 16, covering 37 credited entries across a soft-launch January, a clean February, and the first half of March. A brief cross-model window opened Mar 17 when we brought GPT-5.4 (gpt-5.4-2026-03-05) alongside the GPT-5 baseline; the window ran nine days and produced 19 credited entries before we switched the master flow to Claude Opus 4.6 (claude-opus-4-6) on Mar 26. From Mar 26 through May 13 every credited entry on the published record has run on Claude, 36 trades and plus 11.15R net so far. The story underneath the era split is operational, not competitive. The Q1 model transition was deliberate and fast: nine days of cross-model evaluation, then a decision. The reading is that the architecture did the work and the master analyst executes the playbook; switching the executor was a change in how the playbook is run, not a change in the playbook itself. The Trend Agent's confluence threshold did not move across the cutover. The Macro Agent's gate held. The same continuation-pullback library that paid through February on GPT-5 took the March variance hit and rebuilt across April on Claude without a single threshold change.

Act 1: Jan 12 to end of February, soft launch on GPT-5

SkyAnalyst booted with GPT-5 as the sole master analyst on Jan 12, 2026. Single-model run, no panel, no parallel evaluation yet. January was a soft-launch month with three credited entries across the FTMO and post-FTMO ledger, totalling plus 3.02R. The Jan 12 NAS100 long opened the live record at plus 0.48R, a TP2 print on a Setup 2 breakout-continuation. The Jan 15 NAS100 long added plus 0.78R on a pullback-to-go primary. The month's largest credited single decision landed on the Jan 20 US30 short, a primary fade of the morning bounce that ran the full distance to TP3 for plus 1.76R. Three for three on a thin January sample is not a verdict; it is the opening of the ledger.

February was the first month with real master-automation flow at scale, still entirely on GPT-5. Twenty-four trades. Fifteen winners, nine losers, 62.5 percent win rate, plus 6.66R net. The equity-index complex did the carrying, with US30 producing a string of credited winners across the Feb 9 to Feb 13 cluster and the largest credited single decision of the year so far landing on Feb 11, a US30 long that ran to plus 2.64R. The Feb 24 NAS100 short to TP3 added another credited contribution the following week. February closed cumulative at plus 9.68R YTD on the inception-to-date line. The premise of the playbook held on a single-model run.

Act 2: Mar 17 to Mar 25, the nine-day cross-model window

The first half of March was a continuation of the GPT-5 era. The master flow stayed on GPT-5 through Mar 16, taking the early variance of the month under the launch executor. Cumulative R peaked north of plus 10R in early March, then ground lower through the middle as the tape that paid in February did not pay the same way in March.

On Mar 17 we introduced GPT-5.4 (gpt-5.4-2026-03-05) alongside the existing GPT-5 baseline, opening a brief A/B window across the back half of the variance month. Nineteen credited entries ran through the cross-model period from Mar 17 to Mar 25. The window was not designed as a long bake-off; it was designed as a fast read on whether a newer GPT generation changed the per-trade economics on the same Macro and Cross-Asset gates. After nine days of parallel running, the cross-model period had produced enough sample plus the Claude evaluations had been visible alongside, and on Mar 26 we switched the master flow to Claude Opus 4.6.

The transition was deliberate and fast. Not a long bake-off, a cutover with a clear decision criterion: which executor held threshold math without override patterns through the variance window. The Trend Agent's confluence threshold did not move across the transition. The Macro Agent's regime gate did not loosen. The Risk Agent's 2 percent per-entry sizing did not shift. What changed was the model running the playbook, not the playbook itself.

Act 3: Mar 26 to May 13, settled cadence on Claude Opus 4.6

From Mar 26 forward every credited entry on the published record has run on Claude Opus 4.6 (claude-opus-4-6). The Claude era covers the back week of March, all of April, and May through the thirteenth, 36 trades for plus 11.15R net at 61.1 percent win rate, an average of plus 0.31R per trade.

April was the first full month run entirely on Claude. Eighteen trades, 55.6 percent, plus 2.24R net. The equity-index complex carried it: US30, NAS100, and US500 combined for the bulk of the month's credited R across a clean Claude-only ledger. The lone B-grade entry of the month, the Apr 23 EURUSD short, ran to TP3 for plus 1.67R credited. May through the thirteenth has produced 12 trades for plus 2.50R at 58.3 percent, including a Feb-style Claude run from May 1 to May 6 that printed plus 5.34R across six entries before the May 7 to May 8 USDJPY and NAS100 cluster gave roughly 3R back. The cumulative line has climbed from plus 11.53R at the end of March to plus 15.62R through May 13 without a single window matching March's mid-month dispersion.

What changed across the three eras was not the playbook. The same continuation-pullback library that paid in January and February on GPT-5 kept paying in April and May on Claude. What changed was which executor ran the architecture, and that change happened across a deliberate nine-day window rather than a slow drift. The April and May ledgers are the system's expectancy through a settled cadence under a new executor.

Pattern of the week

The pattern that runs across the year is not a setup; it is a posture, and the posture survived an executor change. The architecture does not consult the recent trade record to size or decline the next decision, and the architecture does not change when the master analyst changes. The Macro Agent grades regime before any pattern. The Trend Agent scores structural confluence and fires only when the threshold clears. The Cross-Asset Agent vetoes false breaks. The Risk Agent enforces fixed 2 percent risk per entry. None of those gates moved when the cumulative line was climbing on GPT-5 in February, none moved when it was grinding lower across the GPT-era variance in early-to-mid March, and none moved when Claude Opus 4.6 took over the master flow on Mar 26.

The pattern visible at year scale

At a single-week horizon, this looks like discipline. At year scale, it looks like consistency: the same threshold math fired across regime rotations from February's clean trend tape on GPT-5 into March's variance window on GPT-5 then GPT-5.4 and back into April's central-tendency month on Claude. The peak-to-trough drawdown of roughly 9.6R between Mar 4 and Mar 24 spanned the tail of the GPT-5 era, the entire GPT-5.4 cross-model window, and the first day of the Claude era. The architecture did not retune through it. The same playbook that paid through the January-and-February run took the March variance hit and rebuilt across April and May without a single threshold change. That is the recurring pattern at year scale: positive expectancy plus posture held equals a cumulative line that climbs slowly through the central tendency and gives back inside the band when the tape turns, regardless of which model is the executor.

Why the same library survived three monthly regimes and an executor change

The continuation-pullback library that produced January's three-for-three open on GPT-5 produced February's bumper on GPT-5, produced March's drawdown across the GPT-5 to GPT-5.4 transition, and produced April's recovery on Claude. The same entry rule fired across all of them. February's tape paid on confluence; March's tape paid on fewer entries and stopped more often on the same threshold; April and May's tape paid again. The reading is that variance inside the central tendency does not require intervention, and an executor change inside a deliberate window does not require playbook intervention either. The system that survives a drawdown and an executor switch without changing threshold math is the system that compounds positively across the next twelve months. That, mechanically, is the year's reason for plus 15.62R YTD across 92 trades.

Decision highlights

The first decision that shaped the year was the choice to open the cross-model window when we did. Through Feb 28 the master flow had run cleanly on GPT-5 and the cumulative line was at plus 9.68R YTD. We waited until the GPT-5 sample had crossed thirty entries on a real flow before introducing a second model on Mar 17. A faster cut would have settled a primary execution path on February's clean trend tape, which is the wrong window to make a long-run model call inside. Holding the single-model run through the first variance week of March produced the sample on which the Mar 26 cutover was made.

The second decision was the cutover itself, made on Mar 25 and executed on Mar 26. Across the Mar 17 to Mar 25 cross-model window the team reviewed GPT-5.4's per-trade R-multiple against the GPT-5 baseline and reviewed Claude Opus 4.6's parallel evaluations against both. After nine days the combined cross-model R-multiple read and posture-discipline read produced a clear-enough call to cut over. From Mar 26 forward every credited entry on the published record has run on Claude. The cross-model panel still runs in parallel evaluation, so the consolidation is reversible if the parallel sample shifts the per-trade R-multiple gap.

The third decision was the one the architecture did not make. Across the Mar 4 to Mar 24 drawdown the cumulative line dropped roughly 9.6R peak-to-trough, spanning the tail of the GPT-5 era, the entire GPT-5.4 cross-model window, and ending the day before Claude took over. The Trend Agent's confluence threshold did not move. The Macro Agent's regime gate did not loosen. The Risk Agent's 2 percent per-entry sizing did not shift. Neither did anything change when the executor swapped on Mar 26. The same playbook that paid through the January-and-February run on GPT-5 took the March variance hit across the GPT transition and rebuilt across April and May on Claude. The decision to hold posture through a drawdown and through an executor change is the decision that compounds positively across the next twelve months.

Loss worth learning from

The cleanest loss-cluster the year produced was the Mar 4 to Mar 24 stretch. Cumulative R peaked north of plus 10R in early March and ground back closer to the Feb-end starting line by the third week, a peak-to-trough drawdown of roughly 9.6R across roughly thirty trading sessions. The cluster spanned two executor eras: the tail of the GPT-5 era from Mar 4 to Mar 16, and the entire GPT-5.4 cross-model window from Mar 17 to Mar 24. No single trade carried the cluster. The cumulative drop came from a string of stops at threshold across the equity-index complex and USDJPY, plus a handful of credited winners that were too small to offset the losers.

What the system saw that was right

A continuation regime that had paid through February on GPT-5 and would pay again in April on Claude. The Trend Agent's confluence cleared threshold on each of the credited entries. The Macro Agent's gate held lean-bull through most of the window. Cross-Asset confirmed the bar-level structural premises were independent. The architecture's reads were the same reads that had paid through February and would pay through April; the tape simply did not deliver across the variance window.

What the tape produced

March's variance ran higher than February's, with more single-day rejections through structural confluence and a higher rate of stops on the same threshold math. The Cross-Asset Agent did not veto more entries because the structural premises remained independent at the bar level. Three correlated stops on the equity indexes early in the third week were the visible turn; the cumulative line bottomed near Mar 24 and rebuilt across the closing partial of the GPT-5.4 window and the opening week of the Claude era.

What we would do the same

Hold posture. The entry rule is the entry rule. The same threshold math that paid through the January-and-February run on GPT-5 took the March variance hit across the GPT-5 tail and the GPT-5.4 window and rebuilt across April and May on Claude. A streak-aware filter would lower realized expectancy across calendar windows, not raise it. The Mar 4 to Mar 24 drawdown was the cost of running the playbook through a regime that did not pay it; the February run on GPT-5 and the April recovery on Claude were the payoff of running the same playbook through regimes that did. Both came from holding posture without changing threshold math through the drawdown, regardless of which executor was running the master flow.

From the desk

Through May 13, 2026, the cumulative ledger reads plus 15.62R YTD across 92 trades from January 12 inception. The same $100,000 simulated account at 2 percent risk per trade sits at $131,242.81 on the static line, where each entry risked a fixed $2,000, and at $134,086.32 on the compounded line, where each entry's risk dollars scaled with the running balance. The two figures diverge by roughly $2,844 across four months and a week. That spread is the cost or benefit, depending on view, of compounding through a positive-expectancy edge: when winners cluster after wins, compounded outpaces static; when winners cluster after losses, static outpaces compounded. The 2026 distribution has favored compounded across the post-cutover rebuild on Claude. That is the disciplined-sizing math working as designed, not a sign one approach beats the other in the abstract.

The honest reading of the year so far is that the system delivered what a positive-expectancy playbook delivers across a soft launch on GPT-5, a clean trend month on GPT-5, a variance test that spanned a GPT-5 to GPT-5.4 to Claude transition, and a settled cadence on Claude. The cumulative line climbed through January and February on GPT-5, gave most of February's run back across mid-to-late March through the cross-model window, and rebuilt steadily across April and May on Claude. The architecture did not change posture across any of the four monthly regimes or across the executor change. The Trend Agent's confluence threshold did not move. The Macro Agent's gate held. The Risk Agent's 2 percent sizing did not shift. Same playbook, four regimes, three executors, plus 15.62R cumulative.

The credited recap underweights live broker fills by design. Every R-multiple here is computed on a TP1 exit for every winner; a subscriber on the published scale-out plan would have closed the year-to-date further into the green because a non-trivial fraction of the 53 credited winners ran past TP1 to TP2 or TP3 in the live broker tape. The plus 15.62R is the floor of the credited projection, not the realized ceiling. The compounded $134,086.32 line is the conservative reading of the year so far.

What carries forward into the rest of 2026 is the same playbook and, for now, the same master analyst. The system does not tune based on the calendar; it tunes based on the rolling 100-trade window. With 92 trades on the inception-to-date ledger, the next eight entries close the first rolling-100 window since launch and are the first sample on which any setup-criteria adjustment can reasonably be made. Until then, the architecture holds posture: same Macro Agent gate, same Trend Agent threshold, same Risk Agent sizing, same Claude Opus 4.6 executor. The next monthly recap closes May; the next published case study lands when the master flow produces a decision worth walking. The cumulative ledger and the published case studies will continue to run inside the same architecture they have run inside since Jan 12.

Monthly views: February 2026 monthly recap, March 2026 monthly recap, April 2026 monthly recap.

What we're tuning

The year-to-date sample has produced two candidate signals worth watching forward into the rest of 2026. USDJPY at minus 0.53R on eight entries (37.5 percent win rate) is the weakest fit on the canonical instrument grid; the question forward is whether the playbook's setup criteria need to tighten on the pair specifically, or whether the eight-entry sample is too small to act on. The rolling-window monitor is set on the pair; we will not retune setup criteria on a sample below 30 entries. The pair's drawdown spans the GPT-5.4 cross-model window and the Claude era, so the read is not executor-specific.

The second candidate is the cross-model parallel sample on the Claude side. The Claude era's 36 entries at plus 0.31R per trade against the combined GPT eras' 56 entries at plus 0.08R per trade is the headline per-trade R-multiple gap, and it is the gap that produced the Mar 26 cutover. If, across the next thirty to forty parallel evaluations, the GPT panel begins closing the per-trade R-multiple gap on the same instrument grid, the master-flow centralization on Claude becomes a candidate for revisit. None of this is an immediate tuning item. The right horizon for any retune decision remains the rolling 100-trade window. The 2026 ledger is now 92 trades; the next eight entries close the first full rolling-100 window since inception and are the first sample on which a tuning call can reasonably be made.